Human-in-the-Loop Task API for AI Agents (REST + MCP)
Give your AI agent real hands: a REST API and an MCP server to create human tasks in 85 countries, collect screenshots, links or text as proof, and approve or reject each one. Works with Claude Code and any MCP client.
By Romain, Founder of Hitworkers · Updated · 8 min read
No subscription · Rejected work is refunded · Real people, with proof
A human-in-the-loop task API lets software, and increasingly AI agents, send a task to a real person and get a verifiable result back. Agents are good at reading, writing and calling tools, but some steps still need a human: searching from a real phone in another country, testing a sign-up flow end to end, recording a voice sample, checking a translation, judging whether a page looks right or publishing from a real account. Hitworkers turns those steps into API calls. Your agent creates a campaign with instructions, a reward and the proof it needs; real workers in 85 countries do the task and send a screenshot, a link, a file or structured fields; the agent reviews each proof and approves or rejects it. Everything is available both as a REST API and as an MCP server, so Claude Code and other MCP clients can use it as a set of tools. Documentation is at /docs, the OpenAPI spec at /openapi.json and a plain-text version for language models at /llms.txt. This guide covers setup, the tools, a complete agent loop, pricing, timing and the guardrails that keep it safe.
Hitworkers vs Generic crowdsourcing without an agent interface
Many crowdsourcing sites were built for humans clicking through a dashboard. Here is what an agent needs, and how Hitworkers provides it.
| Hitworkers | Generic crowdsourcing without an agent interface | |
|---|---|---|
| Interface for agents | MCP server with 18 tools, plus REST and OpenAPI | Web dashboard, sometimes a legacy API |
| Docs readable by LLMs | /llms.txt and /docs.md, plain text | HTML pages written for humans |
| Safe retries | Idempotency-Key on every POST / PATCH | Retries can create duplicates |
| Proof the agent can inspect | Screenshots returned as images to the agent, links, text or fields | Files to download manually |
| Cost control | POST /quote before creating; only approved proofs paid | Prices discovered after launch |
| Human oversight | Our team reviews each campaign before it goes live | Varies |
How it works
Create an API key in your client space and add the MCP server to Claude Code with one command, or call the REST API from your own code.
It reads the catalog, quotes the campaign, then creates it with instructions, reward, countries, pace and the proof required.
It lists proofs waiting for review, inspects screenshots and text, and approves or rejects each one with a reason. Rejected work is refunded.
You approve every submission
Each worker attaches a screenshot or a link. You approve what's valid in one click — rejected work is refunded, so you never pay for a task that misses your brief.
Why advertisers use it
MCP tools, OpenAPI, plain-text docs and idempotent writes: designed to be called by software, not only clicked by people.
Every task returns proof the agent can check before paying, which keeps a human step auditable inside an automated flow.
1,087 activated workers on their own devices, for steps that need a real browser, phone, location or judgment.
Quote first, reserve the budget, pay only approved proofs and stop at any time to recover unused budget.
What human-in-the-loop means for an AI agent
Human-in-the-loop usually describes a person reviewing a model's output. For agents, the loop also runs the other way: the agent delegates a step to a person and reviews the person's output. Typical examples are steps that need a real device or location (search Google from Kenya and screenshot the results), real-world testing (sign up on our staging site from a phone and report what broke), human perception (does this thumbnail read well at small size?), content that must come from real accounts, or data collection a person can do in minutes but a script cannot do reliably. The agent defines the task and the proof, real people execute it, and the agent stays in control by approving or rejecting each result. That last step matters: the agent never has to trust a counter, because each task returns something it can inspect.
- Real-device checks: search results, page rendering or app behaviour from a given country.
- End-to-end tests: sign-ups, checkouts, forms, email confirmation flows.
- Human judgment: readability, translations, first impressions, short surveys.
- Data tasks: collecting or verifying information from public pages.
- Publishing or engagement tasks from real accounts, within platform rules.
Connect Claude Code or any MCP client
Create an API key in your Hitworkers client space; it starts with hw_live_ and is shown only once. In Claude Code, add the MCP server with a single command: claude mcp add --transport http hitworkers https://hitworkers.com/api/mcp --header "Authorization: Bearer hw_live_…" (replace the placeholder with your key). The server uses Streamable HTTP and the same key as the REST API. You can then ask in plain words, for example: launch Google SEO Boost on my pricing page with the keyword 'crm for startups', 100 searches over 5 days, or review the proofs of my last campaign and reject the ones without a screenshot of the comment. Other MCP clients connect to the same endpoint with the same Authorization header. If you prefer plain HTTP, the base URL is https://hitworkers.com/api/v1, requests and responses are JSON, and the rate limit is 120 requests per minute across REST and MCP.
- MCP endpoint: https://hitworkers.com/api/mcp (Streamable HTTP).
- REST base URL: https://hitworkers.com/api/v1.
- Auth: Authorization: Bearer hw_live_… on every request.
- Docs: /docs (human), /docs.md and /llms.txt (plain text), /openapi.json (spec).
- Rate limit: 120 requests per minute, REST and MCP together.
The tools your agent gets
The MCP server exposes 18 tools that mirror the REST endpoints. An agent typically starts with get_catalog and get_account, prices the job with quote_campaign, launches with create_campaign and then works through list_submissions, get_proof_file and review_submission. The table lists the main ones; the full list and parameters are in the docs.
| MCP tool | REST endpoint | What it does |
|---|---|---|
| get_catalog | GET /catalog | Templates, simple templates, audiences, prices and minimums |
| quote_campaign | POST /quote | Exact total and cost per task, nothing created |
| create_campaign | POST /campaigns | Create a campaign or a draft (budget reserved) |
| get_campaign | GET /campaigns/{id} | Status, counts, questions from our team, suggested price |
| campaign_action | POST /campaigns/{id}/actions | pause, resume, stop, add_slots, answer, pace, accept_price |
| list_submissions | GET /campaigns/{id}/submissions | Proofs, e.g. status to_review |
| get_proof_file | GET /submissions/{id}/image | A worker's screenshot or file, returned as an image |
| review_submission | POST /submissions/{id}/review | approve, reject, redo or report a proof |
A complete agent loop
A reliable agent follows the same loop every time. First it reads the catalog and the account balance. Then it quotes the campaign and creates it with an idempotency key, so a network retry never creates a duplicate or charges twice. Our team reviews every campaign before it goes live, usually within a few hours; some templates go live immediately. During that review the agent should poll the campaign: it may be live, rejected with a note and fully refunded, or waiting on a question from our team or a suggested higher reward that the agent can answer or accept. Once workers start, the agent checks proofs every few hours, compares each one with the instructions, downloads screenshots when needed and approves or rejects with a clear reason. Proofs not reviewed within the 24- or 48-hour review window are approved automatically, so an agent that goes silent still settles fairly. When it has enough results, the agent stops the campaign and unused budget returns to the balance.
- get_catalog + get_account.
- quote_campaign, then create_campaign with an idempotency key.
- Poll get_campaign: live, rejected (admin_note), question (answer) or new price (accept_price).
- Every few hours: list_submissions status to_review, inspect, review_submission with a reason.
- campaign_action stop when done; unused budget is returned.
Pricing, timing and limits for agents
Agents pay the same prices as any advertiser. A custom task costs the worker reward, from $0.15, plus a 10% fee, and each extra screenshot adds $0.03; simple templates such as Google SEO Boost ($0.30 per visit) or YouTube Boost (from $0.20 per 2-minute view) have all-in prices listed in the catalog. The Best Workers audience adds $0.30 per worker. The budget is reserved at creation, only approved proofs are charged, and a 402 response with need_usd tells the agent the balance is too low. Timing is the main design constraint: people respond in minutes, not milliseconds. The median worker sends proof 10.1 minutes after taking a task, and how soon the first worker starts depends on countries, reward and pace. Design the agent to create, check back later and continue, not to wait on a single call.
| Item | Value |
|---|---|
| Minimum cost per custom task | $0.165 ($0.15 reward + 10% fee) |
| Extra screenshot | +$0.03 per task |
| Best Workers audience | +$0.30 per worker |
| Median time from task taken to proof | 10.1 minutes |
| Review window before auto-approval | 24 or 48 hours |
| Rate limit | 120 requests per minute (REST + MCP) |
| Approval rate, all reviewed proofs | 93% (13,478 of 14,549) |
Guardrails: what agents may and may not ask people to do
Putting people inside an automated loop comes with responsibilities. Every campaign, whether it comes from the dashboard or an agent, is reviewed by our team before it goes live, and illegal, adult, gambling, misleading or account-selling tasks are rejected. Agents should write numbered steps, one action per step, and state exactly what the proof must show and what will be rejected. Never ask workers for passwords, payment details or personal data beyond what the task strictly needs, and never ask them to write fake reviews. Engagement and traffic tasks remain subject to the platforms' own rules, such as Google's spam policies and YouTube's Fake engagement policy, which apply whether a person or a script performs the action. Finally, keep a human owner for the agent's budget: set a balance you are comfortable with and check the statement regularly.
- Clear, numbered instructions and an explicit proof requirement.
- No passwords, payment data or unnecessary personal data.
- No fake reviews, no misleading or illegal tasks.
- Platform policies apply to human and automated engagement alike.
- A human owns the budget and checks the statement.
Frequently asked questions
What is a human-in-the-loop task API?
An API that lets software or an AI agent send a task to a real person and get a verifiable result back, such as a screenshot, a link, a file or structured answers, which the agent reviews before paying.
How do I connect Hitworkers to Claude Code?
Create an API key in your client space, then run: claude mcp add --transport http hitworkers https://hitworkers.com/api/mcp --header "Authorization: Bearer hw_live_…" with your own key. Claude Code can then create campaigns and review proofs in plain language.
Is there a REST API as well as MCP?
Yes. The REST API at https://hitworkers.com/api/v1 covers the same functions as the MCP server, with JSON requests, idempotency keys and an OpenAPI spec at /openapi.json. Plain-text docs for language models are at /llms.txt.
How much does a human task cost through the API?
The worker reward you set, from $0.15, plus a 10% fee, so from $0.165 per task; +$0.03 per extra screenshot and +$0.30 per worker for Best Workers. Simple templates have all-in prices in the catalog. Only approved proofs are paid.
How fast do humans respond?
The median worker sends proof 10.1 minutes after taking a task. Before that, the campaign is reviewed by our team, usually within a few hours, and the first worker's start time depends on countries, reward and pace. Agents should poll rather than block.
Can the agent see the proof screenshots?
Yes. The get_proof_file MCP tool returns screenshots as images the agent can inspect, and the REST endpoint /submissions/{id}/image returns the file. The agent can then approve, reject, ask for a redo or report the proof.
What tasks are not allowed?
Illegal, adult, gambling, misleading or account-selling tasks are rejected at review, as are requests for passwords or payment details and fake reviews. Engagement tasks remain subject to each platform's own policies.
How many campaigns are created by agents today?
45 of the 519 advertiser campaigns on Hitworkers (about 9%) were created via the API or MCP server as of November 22, 2026. See the Human Micro-Task Benchmark 2026 for the full data.
Related services on Hitworkers
Advertisers running human-in-the-loop api campaigns often combine them with these:
Read the API docs in minutes
Set your price, choose how many people, approve the proof. You only pay for work done right.
Read the API docs →