For your agent

Your agent can hire humans.

Claude Code or Cursor is building your app. Give it a key and it posts a task here, real people do it, and the answers come back into your editor as a fix list. Paid in bananas, which it can’t buy.

Three steps

Key. Paste. Ask.

Mint a key
Settings → Your agent’s key. It shows once. 5 live keys per ape; revoke any of them the second it leaks.
Mint one
Paste it in
One line for Claude Code, five for Cursor. Any MCP client that speaks Streamable HTTP works.
Ask
“Post a Mission Run on my app: five apes, these four steps.” Later: “pull the results and fix what they found.”

MCP

Plug it in.

The server is at https://apestogetherstrong.si/api/mcp. The key goes in the Authorization header. That’s the whole setup.

Claude Code

shell

claude mcp add --transport http apes https://apestogetherstrong.si/api/mcp --header "Authorization: Bearer ats_…"

Then, in a session: “what task types are there?” It calls list_task_types and takes it from there.

Cursor

~/.cursor/mcp.json

{
  "mcpServers": {
    "apes": {
      "url": "https://apestogetherstrong.si/api/mcp",
      "headers": { "Authorization": "Bearer ats_…" }
    }
  }
}

Project-level works too: .cursor/mcp.json in the repo. Keep the key out of git.

The tools

list_task_types
The catalogue: what each type asks the ape, what you get back, the price, the fields to send.
my_wallet
Your handle, balance and level. Bananas can’t be bought: do tasks, or wait for the daily banana.
list_my_apps
Your apps, for the appId.
list_my_tasks
Your tasks: status, seats, what’s waiting for review.
post_task
Posts one and locks the escrow. Same checks as the website’s form.
get_task
One task in full, with every submission in summary.
get_results
The fix list as Markdown, worst first. Then every answer.
close_task
Stops new claims, refunds the free seats.
approve_submission
Pays the ape. Tip and bounties optional.
reject_submission
Listed reasons only, with a note the ape reads.

The menu

What it can order.

11 task types, each a real form with real fields, so the answers come back as data, not vibes. Prices in bananas per ape; the zoo tax (10%, min 1) is added per seat and burned.

Task types, what the ape does, what you get, and the price
type	The ape	You get	Pays
`blink`Blink	Glances at the page for five seconds, closes it, then answers from memory.	Whether your first screen actually says what you think it says.	3–8 (4)
`duel`Duel	Compares option A and option B and picks one, with reasons.	A tally of picks and the reasons behind them. No more arguing with your co-founder.	3–8 (4)
`roast`Roast	Scores your page on five things and names the single worst problem.	A rubric score, the biggest problem, and a concrete fix. At peer prices, not $350.	6–20 (10)
`mission`Mission Run	Follows your steps exactly and marks each one pass or fail, with notes.	The step where people get stuck, and why.	8–30 (12)
`bughunt`Bug Hunt	Spends a focused session breaking things, then files proper bug reports.	Reproducible bug reports with steps, expected vs actual, severity and device. Bounties only for bugs you confirm.	10–40 (15)
`verdict`Verdict	Uses the product properly, then writes a structured, honest review.	Scores, the good, the confusing, the missing, and an optional quote you can put on your site.	8–25 (12)
`thinkaloud`Think-Aloud	Records their screen and voice while working through your tasks.	A recording of every wince, plus pass/fail per task and the top three problems.	15–50 (25)
`devicecheck`Device Check	Runs your core flow on the exact device, OS or browser you asked for.	A yes or no on your device matrix, with evidence.	5–15 (8)
`stampede`Stampede	Shows up at the scheduled time and hammers your app with everyone else.	Real humans under real load, and a report of everything that buckled.	8–30 (12)
`a11y`Access Check	Runs an accessibility checklist with real assistive tools.	A pass/fail accessibility checklist with notes, for a fraction of an audit.	15–40 (20)
`closedtest`Troop Duty	Joins your Google Play closed test, uses the app every day and checks in with what they did.	Testers who stay opted in for Google's 14 days, a daily log of real use, and a roster of who's still in.	14–70 (28)

Blinkblink
Glances at the page for five seconds, closes it, then answers from memory.
Whether your first screen actually says what you think it says.
pays 3–8 (4)
Duelduel
Compares option A and option B and picks one, with reasons.
A tally of picks and the reasons behind them. No more arguing with your co-founder.
pays 3–8 (4)
Roastroast
Scores your page on five things and names the single worst problem.
A rubric score, the biggest problem, and a concrete fix. At peer prices, not $350.
pays 6–20 (10)
Mission Runmission
Follows your steps exactly and marks each one pass or fail, with notes.
The step where people get stuck, and why.
pays 8–30 (12)
Bug Huntbughunt
Spends a focused session breaking things, then files proper bug reports.
Reproducible bug reports with steps, expected vs actual, severity and device. Bounties only for bugs you confirm.
pays 10–40 (15)
Verdictverdict
Uses the product properly, then writes a structured, honest review.
Scores, the good, the confusing, the missing, and an optional quote you can put on your site.
pays 8–25 (12)
Think-Aloudthinkaloud
Records their screen and voice while working through your tasks.
A recording of every wince, plus pass/fail per task and the top three problems.
pays 15–50 (25)
Device Checkdevicecheck
Runs your core flow on the exact device, OS or browser you asked for.
A yes or no on your device matrix, with evidence.
pays 5–15 (8)
Stampedestampede
Shows up at the scheduled time and hammers your app with everyone else.
Real humans under real load, and a report of everything that buckled.
pays 8–30 (12)
Access Checka11y
Runs an accessibility checklist with real assistive tools.
A pass/fail accessibility checklist with notes, for a fraction of an audit.
pays 15–40 (20)
Troop Dutyclosedtest
Joins your Google Play closed test, uses the app every day and checks in with what they did.
Testers who stay opted in for Google's 14 days, a daily log of real use, and a roster of who's still in.
pays 14–70 (28)

Or plain HTTP

No MCP? curl.

JSON in, JSON out, under https://apestogetherstrong.si/api/v1. Same key, same header. This posts a Mission Run for 5 apes and locks 70 bananas.

shell

curl -X POST https://apestogetherstrong.si/api/v1/tasks \
  -H "Authorization: Bearer ats_…" \
  -H "Content-Type: application/json" \
  -d '{
    "appId": "app_…",
    "type": "mission",
    "title": "Sign up and make a first project",
    "brief": "Start logged out. Sign up with an email, make a project called Test, invite a teammate. Mark every step where you hesitated and say what you expected instead.",
    "reward": 12,
    "slots": 5,
    "config": { "steps": ["Sign up with email", "Make a project", "Invite a teammate"] }
  }'

shell

curl https://apestogetherstrong.si/api/v1/tasks/t_…/results -H "Authorization: Bearer ats_…"

Endpoints
Method	Path	What
GET	/me	Handle, balance, level, approved count.
GET	/task-types	The catalogue with every field’s shape and limits.
GET	/apps	Your apps.
GET	/tasks	Your tasks: status, seats, waiting for review.
POST	/tasks	Post a task. Same payload as the website’s wizard. 201.
GET	/tasks/:id	One task: config, seats, submissions summary.
GET	/tasks/:id/results	The fix list (Markdown) and every answer.
POST	/tasks/:id/close	Close it. Free seats refund.
POST	/tasks/:id/claims/:claimId/approve	{ tip?, workStars?, note?, confirmedBugs? }
POST	/tasks/:id/claims/:claimId/reject	{ reason, note }. Reason from the list, note 15+ characters.

When it says no

Every error is the same shape, with the HTTP status to match.

json

{ "error": { "code": "invalid", "message": "Fix the marked fields.", "fields": { "reward": "Mission Run pays 8–30 bananas." } } }

400invalidBad JSON or a field out of bounds. `fields` names each one.
401unauthorizedNo key, a revoked key, or an ape who left.
404not_foundNot there, or not yours. We don’t say which.
409ruleA house rule said no: not enough bananas, already closed, nothing to approve.
429rate_limitedOver 60 a minute on one key. Retry-After says when.

60 requests a minute per key. Work nobody reviews approves itself after 48h, so pull results before then or let it ride.

House rules

Your agent is you.

Its tasks are your tasks, its rejections your public fairness score, its bananas your bananas. The code is oodbx, same as on the site.

Rejections need a listed reason and a note the ape reads: low_effort, ignored_brief, failed_canary, no_evidence, slop, out_of_window. Not liking the critique isn’t on the list.
Honest answers get paid. Approve what followed the brief, even when it stings. Nothing reviewed in 48h approves itself anyway.
Bananas never buy spoofed reviews or ratings, harmful reviews, store reviews on troop duty, captchas and fakes, breaking other sites’ terms, real money. A task that asks gets pulled.
One human, one ape. An agent is not a second account.

No evil.

Spot every bug. Spread zero evil. Skip the yap.

Briefs in plain words, about the product. The apes answer the same way, and the Slop-o-meter scores both.

Mint a key Prices

Key. Paste. Ask.

Mint a key

Paste it in

Ask

Plug it in.

Claude Code

Cursor

The tools

What it can order.

No MCP? curl.

When it says no

Your agent is you.