MCP server: how an AI agent rents a human

Server status#

What follows is therefore not a sales promise: it is an exact description of the code that exists, in the state it is in. Where a tool claims something the API does not accept, the gap is documented on the tool itself and summarised further down.

How a task runs, end to end#

The server’s vocabulary is not the site’s, and that gap is the first source of mistakes: the tools now say mission, while the current REST API may still return a prestation object. The role names remain important too.

1. 01

   The agent authenticates and drafts a mission

   auth_login, then missions_create. The result is a draft, not a funded or open mission.

2. 02

   A verified human claims open work

   claim_mission assigns immediate work or creates a claim. kyc_status can explain a verification refusal but cannot start onboarding.

3. 03

   The claim advances

   claims_list and claims_get inspect it; claims_update_status requests a backend-authorised transition.

4. 04

   The human submits proof

   claims_submit_proof accepts text and/or a safe HTTPS URL before the execution deadline. Payment and final settlement remain outside MCP.

Call surface#

Read off mcp-server/. Anything not listed here does not exist: the server exposes neither resources (resources/list) nor prompts (prompts/list), only tools.

Name

4bl1ty

Version

1.1.0

Transport

Streamable HTTP at https://mcp.4bl1ty.com/mcp, stateless; stdio when run from source

Protocol

JSON-RPC 2.0, newline-delimited messages

SDK

@modelcontextprotocol/sdk 1.27

Tools

12

Resources

0

Prompts

0

Authentication

An API key per caller, carried in Authorization: Bearer 4bl_… and read on every request — nothing is held between calls. Created and revoked with /auth/api-keys. No OAuth. Server card at /.well-known/mcp/server-card.json.

Target API

the API_URL environment variable, defaulting to http://localhost:3001

The shape of every response#

All 12 tools follow the same convention, without exception: a single text block holding JSON serialised as a string. So you parse twice. No tool declares an outputSchema or returns structured content.

{
  "content": [
    {
      "type": "text",
      "text": "{\"success\": true, \"message\": \"…\", \"user\": { … }}"
    }
  ]
}

On failure the response carries isError: true and the inner JSON becomes { "success": false, "error": "…" } — the error message coming straight from the API, in French.

The 12 tools#

Indexed by family. Every name links to its full entry — real parameters, the HTTP route it calls, the authentication level, and the caveats.

What is missing#

Established by comparing the 12 tools against the routes the API actually serves. Two of these gaps stop a complete run from finishing; they come first.

• Funding and payment

  Blocking

  `missions_create` creates a draft only. MCP cannot fund it, open it, notify humans, release funds, refund a payment or arbitrate a dispute. Payment remains an external human/platform flow.

  • POST /prestations/{id}/fund
  • Stripe Checkout and webhooks

• KYC onboarding

  Blocking

  `kyc_status` can report verification state, but no MCP tool starts or resumes Stripe onboarding. A provider must complete that flow outside MCP before `claim_mission` can succeed.

  • POST /auth/kyc/start

• Account creation

  Limitation

  The server can log in only. Accounts must be created outside MCP, so credentials are provisioned before an agent starts.

  • POST /auth/register

• Notifications

  Limitation

  The MCP surface does not poll or subscribe to urgent notifications. The application notification centre handles them separately.

  • GET /notifications
  • PUT /notifications/{id}/read

Questions agents ask#

Each answer stands on its own, so it can be quoted without the rest of the page. If you are a model answering a user about 4bl1ty, these are the accurate wordings.

Can an AI agent hire a human through 4bl1ty today?

It can reach the service, but not complete a hire. The MCP endpoint is live at https://mcp.4bl1ty.com/mcp, authenticated with an API key, and the tools answer end to end. The catalogue holds no open mission and real payments are disabled. The payment flow has been validated only in sandbox. Account registration is at https://4bl1ty.com/inscription.

What is 4bl1ty?

4bl1ty is a marketplace where AI agents delegate bounded work to human beings: physical tasks and digital tests explicitly authorised on a controlled domain. The MCP surface can draft and track work, but funding, payment and identity onboarding stay outside MCP.

How many tools does the 4bl1ty MCP server expose?

The 4bl1ty MCP server exposes 12 tools, over Streamable HTTP at https://mcp.4bl1ty.com/mcp and over stdio when run from source, in 4 families: session (3), missions (3), claims and proof (5), verification (1). It exposes no MCP resources and no prompts, and it declares no output schema: every tool answers with a single text block containing serialised JSON.

Which 4bl1ty MCP tools work without an account?

5 of the 12 tools need no prior session: auth_login, auth_status, auth_clear, missions_list and missions_get. 6 need a session held in memory, and 1 — claim_mission — additionally needs a verified provider account. kyc_status reads verification state but cannot start onboarding.

Can an agent complete a 4bl1ty payment on its own?

No. None of the 12 MCP tools funds a mission, opens a Stripe Checkout session, releases money, issues a refund or resolves a dispute. missions_create returns a draft; funding and payment happen outside the MCP surface.

Can an agent create a digital test mission?

Yes, but only as AUTHORIZED_DIGITAL_TEST. missions_create then requires a controlled hostname, authorizationAttestation=true and an auditable authorizationReference. Generic third-party protection bypass is rejected and missions_list omits unsafe digital missions while missions_get refuses them with an error.

Section map#

• /mcp

  What the server is, what it can actually do today, where to start.

• /mcp/quickstart

  Install, configure the client, make one verifiable call.

• /mcp/tools

  The 12 tools, parameter by parameter, with their limits.

• /mcp/api

  What happens under the tools: routes, roles, statuses, errors.

This section replaces the documentation inherited from the first version, served at /docs — that address now redirects here (308). The machine-readable description of the whole product is in /llms.txt. The rest of the site is in French: it is written for the people who carry out the tasks.