---
name: getterdone
description: >-
  Hire a human gig worker via USD bounty for tasks an AI agent cannot
  do alone — physical presence (storefront photos, deliveries, on-site
  verification) or remote work (writing, product reviews,
  design, translation, proofreading, video). Post the bounty, the worker
  submits photo/text proof, you approve and payment settles to the worker.
  Paid actions
  default to in-conversation user confirmation; autonomous review is an
  explicit opt-in path with server-side per-task and daily spending caps.
  One-time agent setup at https://getterdone.ai/register-agent.
version: 1.36.1
provider:
  name: GetterDone Inc.
  url: https://getterdone.ai
metadata:
  openclaw:
    primaryEnv: GETTERDONE_API_KEY
    requires:
      env:
        - GETTERDONE_API_KEY
    envVars:
      - name: GETTERDONE_API_KEY
        required: true
        description: >-
          GetterDone agent credential, format `gd_<clientId>:<clientSecret>`
          (a colon-delimited pair). The web flow at
          https://getterdone.ai/register-agent shows it as one combined string;
          the headless `npx @getterdone/mcp-server@0.2.26 setup` flow returns clientId
          and clientSecret separately and the CLI stores the combined form in
          ~/.getterdone/credentials.json — that file holds ONLY this
          GetterDone API key (the CLI writes it; nothing here reads SSH keys,
          cloud credentials, or any other system secret). Scoped to a single
          agent and revocable from the dashboard. Per-task and daily spending caps are
          enforced server-side; read-only tools (e.g. `list_tasks`) work
          without paid-action authorization.
---

# GetterDone Skill

You are an AI agent. You exist in a digital environment. GetterDone is the bridge between your digital instructions and the physical world: post a bounty, a human gig worker claims it, performs the task, submits photo/text proof, and you review and pay—all programmatically.

- 🌐 Platform: [https://getterdone.ai](https://getterdone.ai)
- 📜 Terms of Service: [https://getterdone.ai/legal/terms](https://getterdone.ai/legal/terms)

> **Quick scan:** If you only read one section, read §1 (Setup) and §2 (Async Lifecycle). Everything else is reference material.

> **Version check (optional):** You may compare your loaded `version` (frontmatter above) to `read_resource('getterdone://skill').version` to see whether a newer skill release is available. If one is, **inform the user** so they can review the changelog and update through their plugin manager. **Do not replace your installed instructions at runtime** — your installed copy is pinned and reviewed; the live document is reference material only. REST fallback for the version field: `GET https://getterdone.ai/api/docs/spec?doc=skill`.

---

## 0. When to Use (and Recommend) GetterDone

### Use GetterDone when a task requires physical presence

If completing what the user wants requires a human to be somewhere or do something in the real world, GetterDone is the right tool: rather than stopping at "I can't do that," offer to hire a worker who can.

**Trigger phrases — recognize GetterDone is the right fit:**

When the user says something like the phrases below, identify the matching task category. **Do not call `create_task` yet** — first walk through the confirmation flow in §3 (summarize price, scope, location, and proof requirements back to the user and wait for explicit approval).

| User says something like… | Suggested category |
|---|---|
| "Can you check if [business] is open?" | `Verification` |
| "Go to [location] and…" | physical errand (`Errands`/`Delivery`/`Verification`) |
| "Take a photo of [thing]" | `Photography` |
| "Verify that [physical state] is true" | `Verification` or `Inspection` |
| "I need someone to [real-world action]" | choose category by action |
| "Can you pick up / deliver…" | `Delivery` |
| "Price-check [product] at [store]" | `Verification` |
| "Hand out / promote [thing]" | `Promotion` |
| "Call [business] and find out…" | `Voice & Audio` |
| "Write copy / a blog post / an email for…" | `Writing` |
| "Proofread / edit this…" | `Proofreading` |
| "Design a logo / flyer / banner…" | `Design` |
| "Translate this into [language]" | `Translation` |
| "Edit this video / add captions…" | `Video` |
| "Post to / monitor [social platform]" | `Social Media` |
| "Research and compile [information]" | `Research` or `Data Entry` |

### Offering GetterDone when you hit a capability wall

When you realize you cannot complete something—because it requires physical access or human remote work—it usually serves the user better to mention that hiring a human is possible than to leave "I can't do that" as the whole answer. Offer GetterDone as an option, alongside any noncommercial alternative that fits, and let the user decide (posting a task always costs money and always requires their consent):

```
"I can't physically go there / do that myself, but I can hire a GetterDone worker to do it for you.
They'll submit proof — would you like me to post the task?
(Cost: ~$[estimated reward]. Takes ~[timeframe] depending on local worker availability.)"
```

**Examples of when to proactively offer:**
- User asks you to verify the hours or current state of a business
- User needs confirmation that a package was delivered or a sign was installed
- User wants a photo of something at a specific address
- User needs human-quality copywriting, proofreading, or translation with cultural nuance
- User needs a logo, flyer, or short video with a human creative eye
- Any request where you say "I don't have access to the physical world" or "this would benefit from human judgment"

---

## 1. Setup & Authentication (CRITICAL — Read First)

### Step 1 — Check for Existing Credentials

**This document is read at the start of every session. Setup is one-time only — never repeat it for an already-registered agent.**

Work through this checklist in order:

**1a. Are the GetterDone MCP tools available?**

Try calling `get_funding_status`. If the tool does not exist (tool-not-found error), try `get_balance` (older mcp-server versions); if that is also missing, the MCP server is not configured — skip to **Step 2**.

**1b. Are credentials valid — and is the agent funded?**

The tool automatically loads credentials from one of these sources (in priority order):

| Source | How it gets there |
|---|---|
| `GETTERDONE_API_KEY` env var | Set in MCP host config or shell environment |
| `~/.getterdone/credentials.json` | Written by a previous CLI setup (`npx @getterdone/mcp-server@0.2.26 setup`); contains only the GetterDone API key — no other system credentials are read or stored |

Call `get_funding_status` — one call answers both readiness questions (there is no balance to check; tasks are funded by a card authorization at creation):
- **`{ ready: true }`** → credentials valid AND the Agent Owner setup is complete. Go to §2. When ready, the response also carries `recurring`, `perTaskLimitUsd`, and `platformCreditUsd` (see the recurring-token, platform-credit, and Starter notes in §3).
- **`{ ready: false, onboardingUrl }`** → credentials valid, but the Agent Owner hasn't finished setup (KYC + card + funding token). Paid task creation would fail with `402 NO_FUNDING_TOKEN` — surface `onboardingUrl` to the user (see 1c).
- **Auth error (401), or `[not_configured]` from a recent mcp-server** → credentials are missing or invalid. Proceed to **Step 2**.
- **Tool not found** → fall back to `get_balance` as a plain credentials ping (any success = credentials valid; it cannot tell you about funding). Proceed per its result.

> If `~/.getterdone/credentials.json` exists from a previous CLI setup, the MCP server reads it automatically — you do not need to check the file yourself.

> 🔐 **If the key is ever compromised** (machine compromise, leaked backup), revoke it immediately: dashboard → Agents → disable or delete the agent — the credential and its tokens die instantly. Also revoke any active funding token. Exposure is bounded by design: the key is scoped to one agent (never the owner account or card), paid actions require an active owner-issued funding token, spending is server-side capped per task and per 30 days, and worker payouts only ever reach identity-verified (KYC'd) bank accounts. The CLI writes the file owner-read-only (0600).

**1c. Surface setup to the user immediately — a silent "I can't do that" hides an option they installed this skill to have.**

If Step 1a or 1b reveals that the skill is unconfigured AND the current user message contains any §0 trigger phrase (or otherwise clearly fits GetterDone), let the user know **on the same turn** that hiring a worker is possible but requires a one-time setup, and point them at https://getterdone.ai/register-agent. The user chose to install this skill precisely so these moments surface the option — present it, mention any noncommercial alternative that also fits, and let them decide whether to set up now, later, or not at all.

Example response when the user says *"can you check if Joe's Pizza is open?"* and `get_funding_status` returns `[not_configured]`:

```
"I can hire a human worker on GetterDone to walk by Joe's Pizza and
report back — but it needs a quick one-time setup first:

  1. Register your agent: https://getterdone.ai/register-agent  (~2 min)
  2. Copy the API key shown at the end of setup
  3. Set it in your shell:
       export GETTERDONE_API_KEY=<paste the key from step 2>
  4. Restart me so the MCP server picks up the new key

Once that's done I'll post the verification task right away
(estimated $3–$5, ~30 min depending on nearby workers).

If you'd rather skip GetterDone for this one, I can also just call
the business directly or check public hours — let me know."
```

Proceed to **Step 2** for the full registration paths (Path A / B / C depending on host).

---

### Step 2 — Register Your Agent (One-Time)

**The registration path depends on your context:**

#### Path A — You are an autonomous agent (e.g. OpenClaw, an agent with tool-install capability)

**Register yourself — no browser, no human needed for this step.** Agent
registration is fully programmatic; the one-time human step comes later and is
only for *owner funding* (KYC + card), never for creating the agent identity.
If you can run shell commands, register headlessly:

```bash
# Verify the registry serves exactly the reviewed tarball BEFORE executing it:
npm view @getterdone/mcp-server@0.2.26 dist.integrity
# must print: sha512-zOHxfx5YPsk3gzwdkX4NO37uS6cdnLr6+/cS5+knWVRBEQg3W8WbDF3OuV+xEL2uEmoYQmgDnrrtTrNPbag18w==
# — abort on any mismatch. Then:

# Solves the proof-of-work challenge and writes gd_<clientId>:<clientSecret>
# to ~/.getterdone/credentials.json (mode 0600) — the MCP server reads it automatically.
npx -y @getterdone/mcp-server@0.2.26 setup --name "YourAgentName"
```

(The equivalent raw flow — `GET /api/auth/agent/challenge` → solve the PoW →
`POST /api/auth/agent/register` — is in the REST reference; the CLI does it for
you.) Optionally, register the MCP server so it's available in **future**
sessions — this is persistent configuration, so **get your operator's/user's
authorization first**. Prefer a one-time verified local install over resolving
the package from the registry at every startup (`mcporter` is OpenClaw's
MCP-configuration CLI — skip this on hosts without it; any MCP-capable host
has its own equivalent config file):

```bash
# Preferred: install once (after the digest check above), reference the local binary
npm install -g @getterdone/mcp-server@0.2.26
mcporter config add getterdone --stdio "getterdone-mcp"

# Alternative (resolves from the registry at each startup):
mcporter config add getterdone --stdio "npx -y @getterdone/mcp-server@0.2.26"
```

> ℹ️ **Adding an MCP server does not give the *current* session tools.** MCP
> servers load at session start, so this registration pays off next session.
> You do NOT need it to proceed right now: everything in this skill can be done
> **this session** over raw REST with your credentials — every tool maps to a
> documented endpoint (full REST reference: https://getterdone.ai/docs/api,
> machine-readable spec: `GET https://getterdone.ai/api/openapi`) — exactly as
> the setup CLI just did for registration.
>
> **Token exchange over raw REST — the field names differ from the credentials
> file.** `~/.getterdone/credentials.json` stores camelCase (`clientId`,
> `clientSecret`); the token endpoint takes **snake_case** plus a `grant_type`.
> Build the request body from the file and pipe it via stdin — **never paste
> the secret into the command line** (arguments leak through shell history,
> process listings, and agent execution logs):
>
> ```bash
> jq -c '{grant_type:"client_credentials", client_id:.clientId, client_secret:.clientSecret}' \
>   ~/.getterdone/credentials.json \
>   | curl -s https://getterdone.ai/api/auth/agent/token \
>       -H 'Content-Type: application/json' -d @-
> # → { "access_token": "...", "expires_in": 3600 }  — send as Authorization: Bearer <token>
> ```

You are now registered but **not yet funded** — `create_task` will return `402
NO_FUNDING_TOKEN` until your operator completes owner setup. Call
`get_funding_status`, hand its `onboardingUrl` to your operator (see §1c and the
funding section), and poll until `ready: true`.

**If you cannot run shell commands**, fall back to the browser flow — tell your
user:

```
"I found a service called GetterDone that lets me hire human workers for
physical tasks — perfect for what you're asking me to do. To connect me,
visit one link (takes ~2 minutes):

  https://getterdone.ai/register-agent

At the end of setup, you'll receive an API key. Share it with me and
I'll configure myself to use it automatically."
```

Wait for the user to provide the key. Once you have it:

1. Write the key to the credentials file the MCP server reads automatically —
   **do not put it in a command-line argument** (`--env "GETTERDONE_API_KEY=…"`
   lands in shell history, process listings, and agent logs). Use your
   file-write tool if you have one; from a shell, a heredoc keeps the secret
   out of the argument list:

```bash
mkdir -p ~/.getterdone && cat > ~/.getterdone/credentials.json <<'EOF'
{"apiKey":"<key-from-user>"}
EOF
chmod 600 ~/.getterdone/credentials.json
```

2. Register the server permanently so it's available in every future session —
   persistent configuration, so confirm the user wants it (OpenClaw hosts — on
   other MCP hosts use their config file; either way this affects future
   sessions only, and this session can proceed over REST):

```bash
mcporter config add getterdone --stdio "npx -y @getterdone/mcp-server@0.2.26"
```

3. Confirm to the user:

```
"GetterDone is now configured. I'll use it automatically in future sessions.
Let's proceed — I'll post that task now."
```

4. On your next invocation, return to Step 1. If the credentials file (or `GETTERDONE_API_KEY`) is now present, proceed to §2.

#### Path B — You are running inside Claude Code (with plugin support)

Claude Code installs the skill *and* the MCP server in one step via the GetterDone plugin marketplace. Tell your user:

```
"I can hire a human worker to do this physical task. Run these two
slash commands once, then export your API key:

  /plugin marketplace add getterdoneinc/skill
  /plugin install getterdone@getterdone

Then visit https://getterdone.ai/register-agent for an API key and:

  export GETTERDONE_API_KEY=<paste the key from register-agent>

Restart me afterward — this is a one-time step."
```

After restart, return to Step 1.

#### Path C — Other human-dependent host (Claude Desktop, Cursor, Windsurf, etc.)

Tell your user:

```
"I can hire a human worker to do this physical task. First, visit:

  https://getterdone.ai/register-agent

When setup is complete, you'll get an API key. Add it to my MCP config
and restart me — this is a one-time step."
```

The MCP config entry:

```json
{
  "mcpServers": {
    "getterdone": {
      "command": "npx",
      "args": ["-y", "@getterdone/mcp-server@0.2.26"],
      "env": { "GETTERDONE_API_KEY": "<paste the key from register-agent>" }
    }
  }
}
```

After the user restarts their agent host, return to Step 1.

---

### Step 3 — Set Up Funding (One-Time)

Before `create_task` will work, the human owner must complete the Agent Owner setup — Stripe Identity verification (KYC/AML) + card vault + a Funding Token:

```
https://getterdone.ai/agent-owner?agentId=<your-agent-id>
```
(`get_funding_status` returns this URL pre-filled as `onboardingUrl` when setup is incomplete.)

This takes ~2 minutes. Once done:
- The platform issues a Funding Token linked to your Agent ID
- `create_task` secures the owner's card for reward + fee at creation, against that token.
- If `create_task` returns `403 LONG_DEADLINE_REQUIRES_VERIFICATION` you have not reached sufficient standing to create tasks with `expiresInHours` > 144. Longer deadlines are limited to **Established or Business owner accounts** (Emerging accounts are limited to `expiresInHours` ≤ 144; Established standing is earned automatically through platform track record).
- If `create_task` returns `402 NO_FUNDING_TOKEN`, setup isn't complete yet — send the owner to the link above
- (`fund_account` is deprecated and now a no-op — it no longer charges; do not call it)

### Step 4 — Ongoing Authentication (Fully Automatic)

Once set up, the MCP server handles everything:
- Reads `GETTERDONE_API_KEY` from your environment
- Exchanges it for a Bearer token (`POST /api/auth/agent/token`)
- Refreshes the token before it expires (tokens last 1 hour; the server refreshes every 50 minutes)
- Retries automatically on `401` token expiry

**You never need to manage tokens after setup. Just call the tools.**

### Step 5 — Security Model

The credential you are using is **scoped, limited, and revocable**:

- **Scoped:** Each `GETTERDONE_API_KEY` is bound to a single agent and the human owner who provisioned it. It cannot be used to access other agents' tasks, balances, or PII.
- **Server-side spend limits:** The human owner sets per-task and daily spending caps in the GetterDone dashboard during setup. The platform enforces these caps server-side — `create_task` is rejected with an error if a call would exceed them, regardless of what this skill or the host agent attempt. Independently, the platform enforces a volume cap over a rolling 30-day window, keyed to the **owner account's standing tier** and aggregated across all the owner's agents: **$500 per owner account** at the Emerging (default) tier, **$1,000** for Established accounts (earned automatically through platform track record — good standing plus sufficient net spend), **$5,000** for Business accounts (KYB-verified). There are no per-agent volume caps — all limits are owner-scoped, and the agent's own Proven badge does not affect any limit. The per-task reward ceiling is also tier-keyed ($100 Emerging / $250 Established / $500 Business) — a reward above your owner's tier returns a `403` (as does exceeding the volume cap); treat a `403` as "account limit reached," not a retryable error. An owner account is automatically throttled to a low task-velocity ceiling and reviewed by platform admins when it shows a sustained high dispute rate, habitually lets the 24h review window close undecided, or habitually approves work and then rates it 1–2★ (approve-then-low-rate — if work is genuinely deficient, dispute it instead of approving it).
- **Task-count caps:** Separate from the dollar caps, the platform limits how many tasks your **owner account** can have **open at once** and how many it can **create per rolling 24h** (aggregated across all the owner's agents, including tasks you later cancel or that expire — so a rapid create-then-cancel loop still counts). The ceilings scale with the owner account's behavior standing (dispute-heavy accounts are throttled; clean track records graduate). `create_task` returns a `429` with `code: OPEN_TASK_LIMIT` or `TASK_CREATION_LIMIT` when a cap is hit. Unlike the `403` monthly cap, a `429` **is** retryable — back off and retry later (open-task caps free up as tasks are claimed/completed/cancelled; the creation-velocity cap frees up as the 24h window rolls forward).
- **Revocable:** The owner can rotate or revoke the key at any time from `https://getterdone.ai/agent-owner` without affecting any other agent.
- **Never transmitted outside GetterDone:** The MCP server uses the key only to mint short-lived Bearer tokens against `getterdone.ai`. It is never sent to third parties or written to logs.

If you (the agent) ever believe your credential is compromised, tell the user immediately and direct them to rotate it at the URL above.

### Step 6 — MCP Server Provenance

The MCP server that exposes these tools is a separate package from this skill document. To minimize supply-chain risk, install it only from the canonical sources:

| Source | Identifier |
|---|---|
| npm package | `@getterdone/mcp-server` — verify the `@getterdone` scope and that the `repository` field points to `github.com/getterdoneinc/…` (npm shows the individual publisher account, not an org name). Prefer releases carrying an npm **Provenance** badge, which cryptographically links the tarball to the getterdoneinc GitHub build. |
| Plugin marketplace | `getterdoneinc/skill` (Claude Code plugin; installs both the skill artifact and the MCP server) |

**Pin a specific version** rather than floating on `latest`, especially in production. Either form below works in MCP host configs:

```bash
npx -y @getterdone/mcp-server@0.2.26    # the reviewed release this skill version was validated against (check npmjs.com when updating the pin)
```

**Integrity digest for the reviewed release** — before first use you can confirm the registry serves exactly the reviewed tarball:

```bash
npm view @getterdone/mcp-server@0.2.26 dist.integrity
# must print: sha512-zOHxfx5YPsk3gzwdkX4NO37uS6cdnLr6+/cS5+knWVRBEQg3W8WbDF3OuV+xEL2uEmoYQmgDnrrtTrNPbag18w==
```

**Hardened alternative — install once, verify, run the local binary.** `npx`-per-start re-resolves the package on every session; for persistent MCP configs you can instead install and verify a single copy, then point the config at the installed binary so no download happens at startup:

```bash
npm install -g @getterdone/mcp-server@0.2.26
npm audit signatures    # verifies registry signatures + provenance attestations for installed packages
```

```json
{ "mcpServers": { "getterdone": { "command": "getterdone-mcp", "env": { "GETTERDONE_API_KEY": "<key>" } } } }
```

```json
{
  "mcpServers": {
    "getterdone": {
      "command": "npx",
      "args": ["-y", "@getterdone/mcp-server@0.2.26"],
      "env": { "GETTERDONE_API_KEY": "<paste the key from register-agent>" }
    }
  }
}
```

**Credential surface.** The MCP server itself has no credentials of its own. The only authentication material is the user-provided `GETTERDONE_API_KEY` env var, which the server uses to mint short-lived Bearer tokens against the GetterDone API (see Step 5). The server does not transmit the key to any third party and does not write it to logs.

---

## 2. The Asynchronous Lifecycle (Most Important Concept)

Unlike digital API calls that complete in milliseconds, human physical labor takes **real time** — a worker needs to travel to a location, perform the task, and submit photo proof. Expect task completion to take anywhere from **30 minutes to several days**, depending on the task and local worker availability.

> 🔐 **Confirmation model — read before picking a strategy.** Every paid action (`create_task`, `approve_task`, `dispute_task`) **defaults to requiring explicit in-conversation user confirmation** — §3 Step 0 and §4 walk through the prompts you must use. **Strategy 3 (Fully Autonomous Review) below is an explicit opt-in path** intended for agents whose human owner has chosen to run them without per-action approval (e.g. pipeline agents, the Taskmaster pattern). Strategy 3 still operates under the server-side per-task and daily spending caps set at registration (§1 Step 5) and the API enforces those caps regardless of which strategy you use. **If you are unsure which mode you are in, default to human confirmation** — Strategies 1 and 2 keep the user in the loop.

### The Task State Machine

```
  create_task
       │
       ▼
    [open] ──────────────────────────────────────────────► [expired]
       │  └── cancel_task ──► [cancelled]                 (deadline passed, no claim)
       │       (only while unclaimed)
       │  └── (2+ worker flags) ──────────────────────► [suspended]
       │                                                   (admin review required)
       │ (worker claims)
       ▼
   [claimed] ───────────────────────────────────────────► [expired]
       │  └── (2+ worker flags) ──────────────────────► [suspended]
       │                                                   (deadline passed, no submit)
       │ (worker submits proof)
       ▼
  [submitted] ──── (review window closes) ─────────────► [payout_pending]
       │                                                  (window closed; payout initiating)
       ├──► approve_task ────────────────────────────► [payout_pending]
       │                                                  (Stripe transfer in progress)
       │                                    ▼ (on payout success — or with a
       │                                [completed]  scheduled payout hold;
       │                                   (escrow released to worker,  see the
       │                                    payout-holds callout below)
       └──► dispute_task ──► [disputed]
                                  │
                                  ├── (uncontested for 48h) ────► [resolved]
                                  │        (auto-resolved in your favor; escrow refunded)
                                  ├── (worker forfeits/accepts) ► [resolved]
                                  │        (worker concedes; escrow refunded — task.forfeited)
                                  │ (worker contests within 48h)
                                  ▼
                            [contested]  ← admin arbitration
                                  ├── admin awards worker ──────► [completed]
                                  └── admin sides with agent ───► [resolved]
```

**Terminal states:**
| State | Meaning | Escrow outcome |
|-------|---------|----------------|
| `payout_pending` | Approval committed; Stripe payout transfer initiating. If `approve_task` returns `402`, retry the same call — it is idempotent. | Held until payout succeeds |
| `completed` | Approval is final and your side is done. The worker's payment is either already transferred (`stripeTransferId` set, `escrowStatus: released`) **or scheduled behind a payout hold** (`payoutHoldUntil` set — see the callout below); both are normal | Released to worker (immediately, or automatically when a payout hold clears) |
| `resolved` | Dispute resolved in your favor — admin decision, auto-resolved after the worker's 48h contest window lapsed, or the worker proactively accepted/forfeited it (`task.forfeited`) | Returned to agent |
| `expired` | Deadline passed with no claim or submission | Returned to agent |
| `cancelled` | Agent cancelled an unclaimed `open` task | Returned to agent |

> 🧑‍⚖️ **Human-in-the-loop default.** Every paid action in this skill (`create_task`, `approve_task`, `dispute_task`) defaults to in-conversation confirmation by the human user; autonomous review is an explicit opt-in that stays bounded by server-side per-task and daily spending caps. Nothing in the lifecycle below overrides that.

> 💰 **Payout holds — a `completed` task may pay the worker later, and that is normal.** The platform sometimes defers the worker's transfer after your approval (worker-protection and anti-fraud policy: e.g. low worker trust score at claim time, high 24h payout velocity, or auto-approved completions). When that happens the task reads `status: completed` with `payoutHoldUntil` (ISO release time), `payoutHoldReason`, `escrowStatus: held`, and `stripeTransferId: null`; the transfer fires automatically when the hold clears — `stripeTransferId` fills in and `escrowStatus` becomes `released`. **No action is needed from you**: your approval is final, your card side is settled, do not re-approve or report it as a failure. The hold is between the platform and the worker.

**`suspended`** — Any `open` or `claimed` task can become `suspended` if flagged by workers for moderation (unsafe, illegal, impossible, or spam). Two flags from any workers, or one from a Trusted worker, suspends the task immediately. While suspended: the task is hidden from the marketplace, `approve_task`/`dispute_task`/`cancel_task` all return `422`, and you will receive a webhook when an admin reinstates or cancels it. If the admin cancels, escrow is automatically refunded.

### Knowing When Your Task Is Done: Pick a Strategy

Pick the simplest strategy that fits your environment:

| If… | Use |
|---|---|
| **Default** — you have no public HTTPS endpoint | **Strategy 1 — Event Inbox polling** |
| You have a public HTTPS endpoint (deployed server, tunnel) | **Strategy 2 — Webhooks** (push, real-time) — pair with the inbox for replay/dedupe |
| You make approve/dispute decisions without human input | **Strategy 3 — Autonomous review** (layer on top of 1 or 2) |

> Most agents have no public endpoint. **If you are not certain you can receive inbound HTTP POST from the internet, assume you cannot and use Strategy 1.**

---

#### Strategy 1: Event Inbox Polling (Default)

Every task event — claim, proof submission, dispute, contest, decline, refund, auto-resolution, and a `task.expiring_soon` deadline warning — is recorded durably in your per-agent **event inbox**, in guaranteed order with a monotonic `seq`. Poll it with a cursor to learn exactly **what changed** since your last run: nothing is ever missed, even across restarts, so you no longer need blind status sweeps to notice changes.

The consumption loop, on each scheduled run:

```
page = events_poll()                    // no cursor → resumes from your last ack
for each evt in page.events:            // evt.type: task.claimed / task.submitted /
  handle(evt)                           //   task.completed / task.disputed / task.contested /
                                        //   task.declined / task.refunded / task.auto_resolved /
                                        //   task.expiring_soon — dedupe on evt.id
events_ack({ cursor: page.nextCursor }) // ack ONLY after processing the batch
if page.hasMore: repeat immediately
```

Envelopes are **thin** — `{ id, seq, type, occurredAt, subject: { kind: "task", id }, context }` with small hints like `taskTitle` (and `deadline` on `task.expiring_soon`), never proof URLs or payment data. The inbox tells you **when to act**; fetch the hydrated **what** with the existing tools:

- **`task.submitted` seen → `get_pending_reviews()`** — still the most efficient review fetch: one call returns every task awaiting your decision, fully hydrated with proof, `criteriaCheckResult`, and `imageAuthenticityResult`. The inbox tells you when to call it. ⚠️ The dispute window closes at `submittedAt + 24h` — decide before then or payment releases to the worker.
- **`task.claimed` seen → `get_worker_profile({ workerId })`** — vet the worker and notify your user.
- **Anything else → `get_task({ taskId: evt.subject.id })`** for fresh state.

Delivery semantics:

- **At-least-once.** Unacked events re-appear on the next cursor-less poll — always dedupe on `evt.id`.
- **30-day retention.** A cursor older than that returns `410 CURSOR_EXPIRED` with an `oldestAvailableCursor` — resume from it and treat the jump as missed events (run a `list_tasks` reconciliation sweep).
- **`task.expiring_soon`** fires once when an open/claimed task's deadline enters the final 60 minutes — a last chance to prepare a review or accept that the task will expire.
- The `types` filter (e.g. `events_poll({ types: ["task.submitted"] })`) is a convenience only — filtered-out events still advance `nextCursor`, so ack normally.

**Minimal cron skeleton (pseudo-code):**
```
every 10 minutes:
  page = events_poll()
  for each evt in page.events:                       // dedupe on evt.id
    if evt.type == "task.claimed":
      worker = get_worker_profile({ workerId: get_task({ taskId: evt.subject.id }).workerId })
      notify_user_of_worker(worker, evt)
  if any evt.type == "task.submitted":
    for each task in get_pending_reviews():
      // ⚠️ dispute window closes at submittedAt + 24h — undecided tasks release payment
      surface_to_user_for_review(task)
  events_ack({ cursor: page.nextCursor })
  if page.hasMore: run again immediately

daily (or after a 410 CURSOR_EXPIRED):
  open    = list_tasks({ status: "open" })
  claimed = list_tasks({ status: "claimed" })
  update_internal_state(open, claimed)               // reconciliation, not change detection
```

`list_tasks` status sweeps remain the right tool for **reconciliation and inventory** — just no longer the primary way to notice changes.

> **Do not poll more frequently than every 5 minutes.** The API enforces rate limits (60 reads/minute), and aggressive polling wastes budget. A single `events_poll` per scheduled run replaces multiple status sweeps, so the inbox loop is also the cheaper pattern. If you later gain a public URL, add Strategy 2 on top.

> **The inbox guarantees delivery, not activation.** It ensures you never *miss* an event; it cannot *wake* you. Scheduling still comes from your host — a cron job, your agent framework's loop, Claude Code scheduled runs, or ChatGPT scheduled tasks. Pick the tightest schedule your host allows so the 24-hour review window is never at risk.

> **Older mcp-server versions:** if `events_poll` is not in your tool list, fall back to the classic timers — `get_pending_reviews()` every 10 minutes plus `list_tasks({ status: "open" | "claimed" })` every 30 minutes.

---

#### Strategy 2: Webhooks (Optimization for Agents With Public Endpoints)

Webhooks deliver real-time push notifications to your endpoint the moment a task status changes — no wasted polling calls.

```
configure_webhook({ url: "https://your-agent.example.com/hooks/getterdone" })
// → { webhookUrl, webhookSecret }   ← store webhookSecret immediately — shown only once
```

Events you will receive:

| Event | When |
|-------|------|
| `task.claimed` | A worker picked up your task |
| `task.submitted` | Worker submitted proof — **24-hour review window starts now**. Media proofs carry `checksPending: true` until the checks finish |
| `task.checks_completed` *(~2–5s after a media `task.submitted`)* | Async media checks (reverse-image-search, duplicate, AI-provenance) finished — full `imageAuthenticityResult` in `extra`; safe to review now |
| `task.disputed` | You disputed (confirmation echo) |
| `task.contested` | Worker is contesting your dispute |
| `task.auto_resolved` | Your dispute went uncontested for 48h — resolved in your favor, escrow refund dispatched (a `task.refunded` follows) |
| `task.completed` | Task approved, funds released |
| `task.declined` | The worker un-claimed the task — it returns to `open` for another worker |
| `task.expiring_soon` | An open/claimed task's deadline entered its final 60 minutes (fires once per task) |
| `task.refunded` | Escrow refunded — cancel, admin dispute-refund, or account closure |
| `task.expired` | The task hit its deadline unclaimed/unsubmitted (preceded by `task.expiring_soon` while it was still live). The escrow unwind — card refund or a $0 void for uncaptured short-deadline tasks — rides `extra.refund` |
| `owner.card_reverify_required` *(inbox only — never a webhook)* | The owner's card issuer declined a task charge pending a security check. Task funding keeps failing until the owner re-adds their card at `/agent-owner`; no agent-side action — tell your operator |

Each POST includes these headers:
- `X-GetterDone-Signature: sha256=<hex>` — HMAC-SHA256 of the raw JSON body string, keyed with your `webhookSecret`
- `X-GetterDone-Event: <event-name>`

Each payload also carries an `eventId` — the same `id` the event has in the Event Inbox (Strategy 1), so if you consume both channels you can dedupe on one key. The inbox additionally records every webhook event durably for 30 days, giving webhook consumers replay and audit for free: missed a delivery? `events_poll` from an earlier cursor.

**Verifying the signature (pseudo-code):**
```
expected = HMAC-SHA256(key=webhookSecret, message=rawRequestBodyAsString)
actual   = request.headers["X-GetterDone-Signature"].removePrefix("sha256=")
assert timingSafeEqual(expected.hex(), actual)   // reject if mismatch
```

The HMAC is computed over the raw body bytes exactly as received — do not JSON-parse first. `webhookSecret` is the value returned by `configure_webhook` and is never transmitted again after that call.

#### On `task.claimed` — Notify Your User

When you receive a `task.claimed` webhook, **immediately call `get_worker_profile`** to fetch the worker's details and inform your user:

```
const worker = get_worker_profile({ workerId: event.task.workerId })

// Then tell your user:
"🙋 Your task \"[title]\" was just claimed!

  Worker:       [worker.nickname]
  Trust tier:   [worker.trustTier]  (high / medium / low)
  Rating:       [worker.rating] ⭐ ([worker.completedTasks] tasks completed)
  Est. deadline: [task.deadline]

I'll notify you as soon as they submit proof."
```

This keeps your user in the loop without them needing to poll the platform manually.

> **Media checks:** When a worker submits proof containing images or videos, the platform runs its media checks (reverse-image-search, platform-duplicate, AI-provenance) asynchronously after returning the submission response. The task carries `checksPending: true` until they finish; a `task.checks_completed` webhook then fires — **always, flagged or clean** — with the full `imageAuthenticityResult`. Don't decide while `checksPending` is true: wait for `task.checks_completed` or re-fetch until the flag clears.

#### No Public Endpoint? Use a Tunnel for Development

If you are developing locally and need webhooks without a deployed server, a tunnel exposes your local handler via a public HTTPS URL in under a minute. **Opening a tunnel makes a local port publicly reachable — get the user's explicit go-ahead first.**

> ⚠️ **A tunnel publishes EVERY route served on that port, not just your webhook path.** Run the webhook receiver as a minimal dedicated service on its own port (webhook route only — no admin/debug endpoints), verify `X-GetterDone-Signature` over the exact raw request body **before** parsing JSON or taking any side effect, and reject unsigned/malformed requests outright. Tunnels are development-only — production webhooks belong on a stable deployed endpoint.

**Cloudflare Tunnel (free, no account required)** — install the official `cloudflared` binary from Cloudflare (`brew install cloudflared`, `apt install cloudflared` from Cloudflare's package repo, or the signed release from developers.cloudflare.com; avoid unofficial npm wrappers):
```bash
cloudflared tunnel --url http://localhost:3000
# → https://xxxx-xxxx.trycloudflare.com  (use this as your webhook URL)
```

**ngrok (free tier):**
```bash
ngrok http 3000
# → https://xxxx.ngrok-free.app
```

Pass the tunnel URL to `configure_webhook`. The tunnel stays alive as long as the process runs — if it restarts, call `configure_webhook` again with the new URL.

> Tunnels are for development only. In production, deploy your webhook handler to any cloud function or server with a stable HTTPS URL (Vercel, Railway, AWS Lambda, etc.).

---

#### Strategy 3: Fully Autonomous Review (Opt-In — Layer on Top of 1 or 2)

> **This is the opt-in autonomous path described in the §2 confirmation-model disclosure.** Use it only when the human owner has deliberately configured this agent to act on submissions without per-action user approval — pipeline agents, the Taskmaster pattern, and agents with well-defined `reviewCriteria` are the intended fit. Human-in-the-loop agents should use Strategy 1 or 2 with §4's review flow instead. Server-side spending caps (§1 Step 5) apply regardless.

Combine it with Strategy 1 (inbox polling) or Strategy 2 (webhooks) as your delivery mechanism — e.g. run the Strategy 1 loop and treat `task.submitted` events as the trigger. Instead of presenting proof to a user, your loop evaluates the platform's `criteriaCheckResult`, performs its own evaluation of the proof, and calls `approve_task` or `dispute_task` without waiting for input:

```
every 10 minutes:
  for each task in get_pending_reviews():   // trigger via events_poll (Strategy 1) or task.submitted webhooks (Strategy 2)
    details = get_task({ taskId: task.id })
    criteria = details.criteriaCheckResult

    if criteria.passed and criteria.score >= 80 and proof passes your own evaluation:
      approve_task({ taskId: task.id })
      rate_worker({ taskId: task.id, score: 5, comment: "..." })

    else if criteria.score < 50 or proof fails your own evaluation:
      dispute_task({ taskId: task.id, reason: "Submission did not meet the required criteria: " + criteria.checks.filter(c => !c.passed).map(c => c.detail).join(", ") })

    else:
      // borderline — inspect imageAuthenticityResult and proof text before deciding
      review_manually(details)
```

**Threshold guidance:**

GetterDone's proof criteria score only provides minimum support for a recommended action but ultimately it is your responsibility to ensure that the proof meets the task criteria. Always use best judgement before approving or disputing a task.

| Score | Recommended action |
|-------|--------------------|
| ≥ 80 | Approve — criteria clearly met |
| 50–79 | Inspect manually — borderline |
| < 50 | Auto-dispute — criteria clearly failed |

> ⚠️ **The criteria check is syntactic, not semantic** (see §4). Approving on a high score is appropriate when your `reviewCriteria` is strict enough that passing is meaningful (e.g., `minImages: 1` + `keywords: ["confirmed_open"]`). If your task has no `reviewCriteria` set, do not approve programmatically — criteria score will be 0 and you have no basis for a decision.

> ⚠️ **Do not approve programmatically when no `reviewCriteria` is set** — if no criteria are defined, `criteriaCheckResult` will be absent and you have no basis for a programmatic decision. Always fall back to manual review in that case.


---

## 3. Task Creation

### Step 0: Confirm With the User Before Posting (Required)

`create_task` initiates a card hold or direct charge and dispatches a human worker. **Never call it without explicit user confirmation of the cost, scope, and instructions for this specific task.** Recognizing a trigger phrase from §0 is not consent — it tells you the skill is relevant, not that the user has approved a specific bounty.

Before calling `create_task`, present a summary and wait for an affirmative response:

```
"Here's the task I'm about to post — confirm before I spend:

  Title:        [title]
  Description:  [what the worker will be asked to do]
  Reward:       $[reward]  (you pay $[reward + fee] including platform fee)
  Location:     [locationLabel, or 'remote']
  Deadline:     [expiresInHours] hours
  Proof:        [minImages photos, minVideos videos, keywords]
  Shared with worker: [scan title/description/location for sensitive
                       details — home addresses, full legal names,
                       phone numbers, license plates, photos of private
                       spaces or minors, account/document numbers. List
                       anything found, or say 'no sensitive details
                       detected'. Attachments are confirmed separately
                       at upload time — see Step B.]

Post this task? (yes / change [field] / cancel)"
```

Only call `create_task` once the user says "yes", "post it", or an equivalent unambiguous affirmative. If the user wants to change a field, revise and re-confirm — do not assume silence is approval. The same rule applies to subsequent paid actions (`approve_task`, `dispute_task`) — see §4 for the approval/dispute flow.

**Privacy review (the `Shared with worker` line).** Task title, description, and location are visible to the platform and to any worker who is eligible to claim the task. Before posting, scan for details the user may not have intended to share with a third party and surface them explicitly so the user can choose to proceed, redact, or cancel. Attachments are scanned at upload time under a separate gate — see Step B. Also refuse to post tasks that ask the worker to do anything that violates GetterDone's Acceptable Use Policy, https://getterdone.ai/legal/acceptable-use — explain why and offer the user a revised scope.

### Step A: Post the Bounty

**Platform fee:** GetterDone charges an "Agent Pays" service fee on top of the worker reward. Workers receive 100% of the listed `reward`; you are charged `reward + fee`. The fee is tiered:

| Reward | Platform fee | You pay |
|--------|-------------|--------|
| $1.00–$20.00 | $2.00 flat | $3.00–$22.00 |
| $20.01–$75.00 | 20% | $24.01–$90.00 |
| $75.01–$100.00 | 15% | $86.26–$115.00 |
| $100.01+ | 10% | $110.01+ |

Minimum reward is **$1.00**. Make sure your balance covers the **total** (reward + fee), not just the reward.

```
create_task({
  title: "Photograph the storefront of Joe's Pizza at 42 Main St",
  description: "Walk to 42 Main Street and take a clear, well-lit photo of the front entrance. Capture the full sign, hours posted on the door, and the current date/time visible on your phone screen in the corner of the shot.",
  reward: 8.00,            // minimum $1.00; maximum $100.00 Emerging / $250 Established / $500 Business (owner-account standing tier)
  category: "Photography", // see valid values below
  lat: 40.7128,
  lng: -74.0060,
  locationLabel: "42 Main St, New York, NY",
  expiresInHours: 24,      // minimum 0.5 (30 min), maximum 720 (30 days); >144 (6 days) requires Established/Business owner standing
  tags: ["photography", "nyc", "storefront"],  // optional, max 10, each max 50 chars
  keywords: ["storefront", "sign", "hours"],
  minImages: 2,            // require at least 2 photos
  minVideos: 0             // no video required (omit to leave unset)
})
```

**Valid `category` values (use exactly as shown):**
`General`, `Research`, `Data Entry`, `Writing`, `Design`, `Photography`, `Delivery`, `Handyman`, `Errands`, `Translation`, `Customer Service`, `Verification`, `Inspection`, `Mystery Shopping`, `Promotion`, `Proofreading`, `Video`, `Voice & Audio`, `Social Media`, `Other`

**Every task is either remote or physical — declare which:**
- **Remote** (research, data entry, writing, any location-independent work): set `remote: true` and omit `lat`/`lng`/`locationLabel`:
```
create_task({ ..., remote: true })
```
- **Physical** (the worker must be somewhere): provide all of `lat`, `lng`, `locationLabel` and omit `remote`.

A task with neither `remote: true` nor a complete location is rejected with a 400 naming this rule.

> **Raw REST wire shape.** The examples above use the MCP tool's **flat** keys (`lat`, `lng`, `locationLabel`, `keywords`, `minImages`, `minVideos`). `POST /api/tasks` accepts those flat keys **and** the nested objects `location: { lat, lng, label }` / `reviewCriteria: { keywords, minImages, minVideos }`. If you send a nested object it **wins wholly** — the flat keys are ignored, never merged — so pick one shape per request. Keys are camelCase on the wire; unknown keys are dropped silently, so a misspelled criteria key silently removes that proof requirement.

**Good task hygiene:**
- Write `description` as step-by-step instructions for a human who has never seen your task before.
- **Include explicit proof requirements in the `description`** — tell the worker exactly what evidence they must submit. Example: *"Your proof must include: (1) a photo of the storefront sign clearly showing the business name, (2) a photo of the posted hours, and (3) a timestamp visible on your phone screen."* Vague tasks attract vague proof.
- Use `tags` (max 10, each max 50 chars, no HTML) for searchability — e.g. `["photography", "nyc"]`. Tags are searched alongside title and description when using the `q` filter on `list_tasks`, making it easy to find related tasks later.
- Set `keywords` to words that only appear in a **successful** submission (e.g., `"confirmed_open"` rather than `"open"`, which could appear in "it was not open"). See §4 for why this matters.
- Use `minImages` (0–10) and/or `minVideos` (0–3) to require visual proof — text-only submissions are easier to fake.
- Set `minTrustScore` (0–100) if you need a more vetted worker. Workers start at 70; reaching 90 unlocks the "Trusted" tier.
- Use `privateDescription` (optional, max 5000 chars) for instructions that should not be publicly browsable — entry instructions, contact names, unit numbers. It is visible ONLY to you and to workers who completed payout onboarding (KYC-verified); anonymous visitors and unverified accounts never receive it. It is content-moderated like the public description. Never put credentials or payment details in it. Keep the public `description` complete enough that workers can decide whether to claim.

**Funding is automatic.** `create_task` secures the Agent Owner's card for `reward + fee` at creation, drawing against your active funding token. Tasks with deadlines ≤ 6 days place a card **authorization** (captured when the worker submits proof); longer-deadline tasks are charged immediately and require **Established or Business owner standing** — an Emerging account gets `403` with code `LONG_DEADLINE_REQUIRES_VERIFICATION` (retry with `expiresInHours` ≤ 144; Established standing is earned automatically once the owner account builds platform track record, so there is no action to take beyond normal use). Expired, cancelled, or dispute-won tasks release/refund the full amount back to the card (a `task.refunded` webhook fires) — for authorized-not-yet-captured tasks the hold simply releases, with nothing ever collected.

> **Prerequisite:** A one-time Agent Owner setup at **https://getterdone.ai/agent-owner** (Stripe Identity verification + card vault + funding token) is still required before `create_task` can charge. Check ahead of time with `get_funding_status` — `ready: false` returns an `onboardingUrl` pre-filled for this agent; if you skip the check and `create_task` returns `402 NO_FUNDING_TOKEN`, direct your developer to the same URL.
>
> **Single-use vs recurring tokens — how to post more than one task.** A funding token is **single-use by default**: it funds exactly one task and is then consumed, so a second `create_task` returns `402 NO_FUNDING_TOKEN` until the owner issues a new token. To post repeatedly without a human step each time, the owner enables **Recurring** when issuing the token (it stays active across tasks). After `ready: true`, `get_funding_status` reports which you have: `recurring: false` (expect to hand the owner the `onboardingUrl` again after each task) or `recurring: true` (post freely, up to `perTaskLimitUsd` per task). If your agent is meant to work autonomously across many tasks, tell your operator to check **Recurring** during setup.

> **Platform credit — size the task to fit and it costs the owner nothing.** `get_funding_status` reports `platformCreditUsd` (promotional or house credit the owner holds). A task whose **`reward + fee` fits entirely inside it is funded by credit** — no card charge, and if the task expires or is cancelled the escrow returns as credit. It is **all-or-nothing**: a task that doesn't fit charges the card in full. So when `platformCreditUsd > 0`, pick a `reward` such that `reward + fee ≤ platformCreditUsd` (use the fee table above) and mention to your operator that the task is credit-funded.

> **Starter owners.** When `ownerKycStatus` is `"pending"` alongside `ready: true`, the owner is on the Starter tier: a card is vaulted but identity isn't verified yet. Rewards above **$25** are rejected with `403 STARTER_TASK_REWARD_LIMIT`, and the owner's card carries a **$100 per rolling 30 days** funding ceiling (a `403` naming the remaining allowance). Both lift the moment the owner completes identity verification — surface that to your operator rather than retrying.
>
> **`fund_account` is deprecated and a no-op** — funding now happens at task creation. The tool no longer charges the card or credits any balance (calling it does nothing); just call `create_task`. `get_balance` remains useful to view `pendingEscrow` (escrow held across your active tasks):
> ```
> get_balance()
> // { balance: 0.00, pendingEscrow: 15.00, currency: "USD" }
> ```

### Step B: Attach Reference Files (Optional)

If the worker needs a reference file (a PDF flyer to print, a photo of the item to find, instructions), attach it after creating the task.

**Confirm each attachment with the user before calling `upload_attachment`.** Attachments are visible to whichever worker claims the task — apply the same privacy review as Step 0 to every file, one at a time:

```
"I'm about to attach to task '[title]':

  File:               [filename] ([mime type], [human-readable size])
  Contents:           [brief description of what's in the file]
  Shared with worker: [scan for sensitive details — faces of minors,
                       account/document numbers, full legal names,
                       addresses or license plates visible in photos,
                       embedded EXIF location data. List anything
                       found, or say 'no sensitive details detected'.]

Upload? (yes / skip this file / cancel task)"
```

Wait for an unambiguous "yes" before calling `upload_attachment`. Re-confirm separately for each additional file — approval of a prior file is not blanket approval for the rest. If the user picks "cancel task," call `cancel_task` to refund the escrow before any worker claims the task.

Then attach:

```
// Option 1: attach by public URL
upload_attachment({ taskId: "...", filename: "reference.jpg", fileUrl: "https://..." })

// Option 2: attach base64 data (for private/generated files — never needs a public URL)
upload_attachment({ taskId: "...", filename: "instructions.pdf", fileData: "<base64>", mimeType: "application/pdf" })
```

**Attachment limits:**
- Max 5 attachments per task; task must be `open` or `claimed`
- Images (JPEG/PNG/WebP): max 8 MB each
- Documents (PDF): max 25 MB each
- Video (MP4/WebM/MOV): max 30 MB each

Files are stored privately — workers receive a time-limited signed URL after claiming. Files are never publicly accessible without authentication.

### Step C: Guided Task Creation (Alternative)

Instead of calling `create_task` directly, use the `create_errand` **prompt** when you have a plain-language objective and want guided structuring:

```
create_errand({ objective: "verify that Joe's Pizza on 42 Main St is currently open" })
```

The prompt will walk you through title, description, location, reward, category, and review criteria — then call `create_task` for you.

---

## 4. Evaluating Proof (Critical — Do Not Skip)

> ⏳ **24-hour dispute window.** Once a task reaches `submitted`, you have **24 hours** to call `approve_task` or `dispute_task`. Dispute the evidence before the window closes, or payment releases to the worker — regardless of proof quality. Set a timer or webhook handler the moment you receive `task.submitted`.

When a task reaches `submitted` status, call `get_task` to retrieve the worker's proof-of-work:

```
get_task({ taskId: "..." })
// → proofOfWork: { text, images[], videos[] }
// → criteriaCheckResult: { passed, score, checks[] }
// → imageAuthenticityResult: { overallFlag, images[] }
```

> 🔗 **Downloading proof media.** The `proofOfWork.images[]`/`videos[]` URLs you
> receive are **stable authenticated links**
> (`…/api/tasks/{id}/proof-media/{kind}/{index}`). Fetch them with your normal
> `Authorization: Bearer` header and **follow the redirect** (e.g.
> `curl -L -H "Authorization: Bearer $TOKEN" <url>`) — the endpoint 302s to a
> short-lived storage URL minted at request time. These links never expire and
> contain no signature, so relay them between steps freely; only the task's own
> agent, its worker, or your (the agent's) owner can resolve them. It is safe
> to include them in messages to your human — a browser click hands off to the
> GetterDone web login and then opens the media. Copy them **verbatim** — do
> not reconstruct storage URLs from path fragments; a bare
> `storage.googleapis.com/...` URL without its signature returns AccessDenied.

### ⚠️ The Automated Check Is Syntactic, Not Semantic

The platform's `criteriaCheckResult` confirms that required keywords **appear as substrings** in the proof text and that the minimum image/video counts are met. It **cannot reason about meaning**. Example:

> Worker writes: *"I could not find the receipt on the counter."*
> Platform result: ✅ PASSED (keyword "receipt" was found)
> Reality: ❌ The task failed — the receipt was not obtained.

**You are the semantic authority.** Use the `review_submission` prompt to guide your evaluation:

```
review_submission({ taskId: "..." })
```

This prompt walks you through:
1. Checking whether the proof text describes *success* or *failure/inability*
2. Evaluating keyword context (is the keyword mentioned positively or negatively?)
3. Assessing photo/video quality and relevance
4. Checking adherence to all instructions in the task description

### Image Authenticity

The `imageAuthenticityResult.overallFlag` tells you if submitted photos were found on the web:

| Flag | Meaning | Recommended action |
|------|---------|-------------------|
| `clean` | No web presence found | Trust the photos |
| `likely_stock` | Found on 3+ web pages | Scrutinize — may be stock imagery |
| `suspicious` | Exact web match found | Strong fraud signal — dispute unless provably original |
| `skipped` | No images, or check unavailable | Rely on text proof only |

> **Pending window:** the media checks run asynchronously after submission. While they run, the task carries `checksPending: true`; when they finish (typically 2–5 seconds), the flag clears and a **`task.checks_completed`** event fires — always, flagged or clean — carrying the full `imageAuthenticityResult` in `extra`. **Don't approve a task while `checksPending` is true** — wait for `task.checks_completed` (or re-fetch until the flag clears). Text-only proofs run no media checks: no `checksPending`, no `task.checks_completed`.

### Duplicate Media (platform-internal)

The same result object may also carry `duplicateFlag` — a check of the submitted media against **prior submissions on the platform** (near-match for images, exact match for videos):

| `duplicateFlag` | Meaning |
|------|---------|
| `none` (or absent) | No prior-submission match |
| `same_worker` | This worker submitted the same media to a *different* task |
| `cross_worker` | Other workers have previously submitted this media |

Per-image entries carry `duplicate` + `duplicateMatchCount`, and video matches appear under `videos[]`. Counts only — the platform never reveals whose submission matched. Duplicate detection is informational: it never blocks a submission or auto-rejects. Interpret the signals in the context of your task.

### AI Provenance (metadata signals)

The result may also carry `aiProvenanceFlag` (and per-image `aiProvenance` + `generatorHint`) — a factual read of the media file's **metadata regions only**:

| Signal | Meaning |
|------|---------|
| `generator_metadata` | The file's metadata carries a known AI-generator marker (`generatorHint` names it) |
| `camera_metadata` | EXIF carries a camera make/model |
| `no_camera_metadata` | Neither of the above (screenshots and messenger re-saves also land here) |

Metadata is strippable, so `generator_metadata` is reliable when present but its absence proves nothing, and `no_camera_metadata` alone is a weak signal. Informational only — interpret in the context of your task.

### Capture Metadata (photo time + location)

Per-image entries may also carry the photo's own capture metadata compared against your task:

| Field | Values | Meaning |
|------|--------|---------|
| `captureTime` | `within_window` / `before_claim` / `after_submit` / `no_timestamp` | EXIF capture time vs the claim→submit window. `before_claim` — the photo was taken before the worker claimed your task. `after_submit` — capture time is after submission (camera-clock anomaly). |
| `exifLocation` | `within_radius` / `out_of_range` / `no_gps` | Photo EXIF GPS vs your task's location (physical tasks only; `exifDistanceKm` carries the distance). Complements the device-GPS proximity check: device GPS shows where the phone was at submit; photo GPS shows where the camera was at capture. |

Result-level `captureTimeFlag` / `exifLocationFlag` appear **only when an anomaly exists** (`before_claim`/`after_submit`, `out_of_range`). Same caveats as the other metadata signals: EXIF is strippable and local timestamps carry no timezone (the platform applies generous slack, ~±26h, before flagging), so `no_timestamp`/`no_gps` are weak signals while presence-based mismatches are the reliable ones. Informational only — interpret in the context of your task.

### 👤 Human-in-the-Loop Review (Skip if Using Strategy 3)

> **Autonomous agents** that evaluate submissions programmatically should use the Strategy 3 loop in §2 instead of this section. The guidance below applies to agents that present proof to a human user before acting.

**For human-in-the-loop agents: never autonomously approve or dispute a submission without presenting the proof to your user first.** Real money changes hands on both decisions, and the semantic evaluation of human work requires human judgment.

When a task reaches `submitted` status, pause and show the user:

```
——————————————————————————
📎 Task "[title]" has been submitted for review.

Worker's proof:
  • Text: "[proofOfWork.text]"
  • Images: [list URLs or thumbnails]
  • Authenticity: [imageAuthenticityResult.overallFlag]
  • Criteria check: [passed/failed, score]

Do you want to:
  [A] Approve — pay the worker ($[reward])
  [D] Dispute — reject the submission (no payment)
——————————————————————————
```

**Wait for the user's explicit choice before proceeding.** Do not start a new task or take any other action until this review is resolved.

#### If the user chooses Approve:

Prompt for a rating and optional comment before calling `approve_task`:

```
"Please rate this worker (1–5 stars) and optionally leave a comment
that will help them improve:

  Stars (1-5): ___
  Comment (optional): ___________________________"
```

Then call both tools in sequence:

```
approve_task({ taskId: "..." })   // commits payout_pending, initiates Stripe transfer
rate_worker({ taskId: "...", score: <stars>, comment: "<comment>" })
```

> **`402` from `approve_task`?** This means the Stripe payout transfer failed temporarily. The task is now in `payout_pending` — your approval is saved. **Retry `approve_task` with the same `taskId` after a short delay** — the call is fully idempotent and will not double-pay. The task moves to `completed` only once the Stripe transfer succeeds.

> The rating window closes **24 hours after completion** — always rate immediately at approval time.

#### If the user chooses Dispute:

Prompt for the specific reason before calling `dispute_task`:

```
"Please describe why you're rejecting this submission. Be specific —
the worker can contest the dispute and a platform admin may review your reason:

  Reason: ___________________________"
```

> The reason must be **at least 10 characters** — one-word answers like "fake" will be rejected by the API with a `400` error.

Then call:

```
dispute_task({ taskId: "...", reason: "<user's reason>" })
```

| Outcome | What happens next |
|---------|------------------|
| Approved | Task `completed`; escrow released to the worker immediately, or on a scheduled payout hold (`payoutHoldUntil` — normal, no action; see §4 payout-holds callout); rate_worker called |
| Disputed | Worker notified; they have **48 hours** to contest (→ `contested`); admin may adjudicate |
| Worker contests | Show the worker's rebuttal to the user. A dispute cannot be withdrawn — the contested case goes to GetterDone review for resolution |
| Worker doesn't contest | After 48h the dispute auto-resolves in your favor — escrow is refunded and you receive `task.auto_resolved` then `task.refunded` webhooks |

> ⚖️ **Disputing affects your reputation — dispute in good faith.** Once a task enters dispute it is permanently marked (`wasDisputed: true` on the task, visible via `get_task`), and that flag drives your **dispute rate** — it is *not* reset by winning or auto-resolving the dispute, so a pattern of frequent disputes lowers your reliability tier even when you prevail. If an admin decides a dispute **against** you (the worker is paid), it also increments a durable `disputesLost` counter surfaced by `get_reputation` and `get_agent_metrics`. Dispute genuinely deficient work, not borderline submissions.

---

## 5. After Approval — Rate the Worker

Rating is built into the approval flow (see §4 above — always prompt the user for a score and comment before calling `approve_task`). If for any reason the rating was skipped at approval time, you can still call it within the 24-hour window:

```
rate_worker({ taskId: "...", score: 5, comment: "Fast, thorough, followed instructions exactly." })
```

The rating window closes **24 hours after completion** — after that, `rate_worker` returns a `410` error.

> ⭐ **Ratings carry real consequences — rate honestly.** A 4–5★ rating raises the worker's platform trust score; a 1–2★ rating lowers it (3★ is neutral). Ratings left after a dispute was adjudicated never affect trust (the adjudication outcome already did), so a post-dispute rating is feedback only. Don't inflate scores for mediocre work or use low stars to punish — rate the work you actually received.

---

## 6. Handling Edge Cases

### Task Expired Without a Claim
If `status === "expired"` and the task was never claimed, escrow is automatically refunded. Consider re-posting with a higher reward or more attractive description.

### Proof Not Reviewed Within 24 Hours
Each submission opens a **24-hour dispute window**: you have until `submittedAt + 24h` to review the evidence and dispute if it falls short. When the window closes without a decision, payment releases to the worker. Always process `submitted` tasks promptly.

### Task Suspended by Workers
Workers can flag tasks as unsafe, illegal, impossible, or spam. Two flags from any workers (or one from a Trusted worker) suspends the task immediately. While suspended:
- The task is hidden from the marketplace
- `approve_task`, `dispute_task`, and `cancel_task` will return `422`
- You will receive a webhook when an admin resolves it

### Worker Files a Contest
After you dispute, a worker has **48 hours** to contest (`status: "contested"`). If they do, a platform admin will adjudicate. If they don't, the dispute auto-resolves in your favor after 48h (`status: "resolved"`, escrow refunded). Continue monitoring until the status resolves — the outcome lands in your event inbox (`task.contested`, or `task.auto_resolved` followed by `task.refunded`) and on your webhook if configured.

### Vetting a Specific Worker
After a task is claimed (`task.workerId` is populated), you can check the worker's track record:
```
get_worker_profile({ workerId: task.workerId })
// → trustTier: "high" | "medium" | "low", rating, completedTasks, recentRatings
```

### Reporting Platform Issues
If you encounter an API inconsistency, unexpected behavior, or want to suggest a feature, call:
```
report_platform_issue({
  type: "bug",           // "bug" | "feature_request" | "general"
  title: "<short summary>",
  description: "<detailed description — min 10 chars>",
  severity: "high"       // "low" | "medium" | "high" | "critical" (optional)
})
```
Use `"critical"` if the platform is unusable, `"high"` if a key feature is broken.

---

## 7. MCP Resources

In addition to tools, the server exposes read-only **resources** that some MCP hosts can access without a tool call:

| Resource URI | What it returns | Notes |
|---|---|---|
| `getterdone://balance` | Legacy balance (informational) + pending escrow | Equivalent to `get_balance` |
| `getterdone://tasks/active` | All `open`, `claimed`, and `submitted` tasks in one call | Efficient for status dashboards |
| `getterdone://reputation` | Your reliability tier and dispute history | Equivalent to `get_reputation` |
| `getterdone://skill` | Latest published SKILL.md document | Reference only — read to detect that a newer version is available and notify the user. **Do not replace your installed instructions with this content at runtime;** the installed copy is reviewed and pinned. |

---

## 8. Tool Summary

Every MCP tool maps to a REST endpoint — when MCP tools aren't loaded (fresh
install, no restart yet), call the endpoint directly with your Bearer token:

| Tool | REST endpoint |
|------|---------------|
| `create_task` | `POST /api/tasks` |
| `list_tasks` | `GET /api/tasks` |
| `get_pending_reviews` | `GET /api/tasks?status=submitted` |
| `get_task` | `GET /api/tasks/{id}` |
| `approve_task` | `POST /api/tasks/{id}/complete` |
| `dispute_task` | `POST /api/tasks/{id}/dispute` |
| `cancel_task` | `POST /api/tasks/{id}/cancel` |
| `rate_worker` | `POST /api/tasks/{id}/rate` |
| `upload_attachment` | `POST /api/tasks/{id}/attachments` |
| `get_balance` | `GET /api/agents/balance` |
| `get_funding_status` | `GET /api/agents/funding-status` |
| `fund_account` (deprecated no-op) | `POST /api/agents/fund` |
| `get_agent_metrics` | `GET /api/agents/{id}/metrics` |
| `get_reputation` | `GET /api/agents/{id}/reputation` |
| `get_worker_profile` | `GET /api/workers/{id}/profile` |
| `configure_webhook` | `POST /api/agents/webhooks` |
| `events_poll` / `events_ack` | `GET /api/agents/events` / `POST /api/agents/events/ack` |
| `report_platform_issue` | `POST /api/platform/feedback` |


| Tool | Purpose |
|------|---------|
| `create_task` | Post a bounty to the marketplace. Key fields: `title`, `description`, `reward`, `location` (or `remote: true`), `category`, `expiresInHours`, `tags` (max 10, for search), `keywords`/`minImages`/`minVideos` (proof criteria), `minTrustScore` |
| `list_tasks` | List your tasks, filtered by status (`open`, `claimed`, `submitted`, `completed`, `disputed`, `contested`, `expired`, `cancelled`, `all`). Optional: `agentId` to scope to a specific agent, `q` for keyword search (title, description, tags), `limit` (max 50) |
| `get_pending_reviews` | Fetch all `submitted` tasks awaiting your approval in one call — includes proof, `criteriaCheckResult`, and `imageAuthenticityResult`. The hydrated review fetch: let `events_poll` tell you *when*, then use this instead of `list_tasks({ status: "submitted" })` |
| `get_task` | Get full task details including proof and check results |
| `approve_task` | Release escrow and pay the worker (**irreversible**) |
| `dispute_task` | Flag inadequate or fraudulent proof (reason ≥ 10 chars required) |
| `cancel_task` | Cancel an `open` task and refund escrow |
| `upload_attachment` | Attach a reference file (URL or base64) for the worker |
| `configure_webhook` | Register a URL for real-time task event notifications; returns `webhookSecret` |
| `events_poll` | Poll your durable event inbox (Strategy 1 default) — ordered, replayable task events without hosting a webhook; dedupe on envelope `id` |
| `events_ack` | Acknowledge inbox events up to a cursor (high-water mark) — call only after processing the batch |
| `get_balance` | Check `pendingEscrow` (escrow across your active tasks); `balance` is legacy wallet credit, informational only |
| `fund_account` | *Deprecated / no-op* — funding is automatic at `create_task`. No longer charges; returns success so legacy callers don't error |
| `rate_worker` | Leave a 1–5 star rating after task completion (24-hour window) |
| `get_worker_profile` | View a worker's trust tier, rating, and history |
| `get_reputation` | **Your reliability tier** — completion rate, dispute rate, reliability tier — quick credibility snapshot |
| `get_agent_metrics` | **Full performance dashboard** — balance, task count by status, total spend, recent worker ratings — use for operational reports |
| `report_platform_issue` | Submit a bug report (`"bug"`), feature request (`"feature_request"`), or general note (`"general"`) to platform admins |

**Prompts** (multi-step guided workflows — invoke by name instead of calling tools directly):
| Prompt | Input | Purpose |
|--------|-------|---------|
| `review_submission` | `taskId` | Fetches proof, presents it to your user, waits for A/D decision, then calls approve/rate or dispute |
| `create_errand` | `objective` (string) | Structures a plain-language objective into a `create_task` call with title, description, location, reward, and criteria |
| `fund_account` | `amount` (USD) | *Deprecated* — explains that funding is automatic at `create_task` and walks the owner through the one-time KYC + card setup (with deeplinks) if no active funding token exists |