OutworxOutworx Docs
Model Context Protocol · Pro and Business

Every project ships with an MCP server.

One URL, one token, twelve tools. Plug your API docs straight into Claude Desktop, Cursor, Cline, or any MCP-aware tool — models can list endpoints, generate typed clients, run real calls, and verify responses against the spec, all without you copy-pasting a single snippet.

Get an MCP serverConnect to Claude Desktop
api/mcp/<slug>
discoverycodegenlive opsQ&A · audit
per-project token

What is MCP

An open standard for AI assistants to call your tools.

Where every AI tool used to need a custom integration, Model Context Protocol speaks one shared protocol — and every MCP-aware client (Claude, Cursor, Cline, Continue) can use it.

Outworx already has a parsed spec, generated TypeScript types, and a hosted Try-It engine. MCP is the natural fit: every capability becomes a tool the model can invoke.

01

The model wants to operate your API

A user asks Claude or Cursor to do something with your API. Without MCP, the model can only describe what to do; with MCP, it can actually do it.

02

Outworx exposes 12 tools

Discovery (list / search / get), code generation (types / SDK / examples), live ops (execute_endpoint, response_schema), Q&A (ask_docs), and audit (recent_calls, get_activity).

03

Calls go through your token

execute_endpoint runs with the user's own credentials — Outworx never sees them. Per-tool permission toggles let you keep destructive operations out of the model's reach.

04

Responses sample the schema

Every live call captures the response shape, infers a JSON schema, and merges it into the stored types. Generated TypeScript improves automatically.

Connect in 30 seconds

Pick your IDE. Paste the snippet.

The dashboard's MCP tab generates a one-click connection snippet for every supported client. Paste it into the config and your AI assistant has 12 new tools.

Claude Desktop
claude_desktop_config.json
{
  "mcpServers": {
    "outworx-acme": {
      "url": "https://docs.outworx.io/api/mcp/acme",
      "headers": {
        "Authorization": "Bearer <token>"
      }
    }
  }
}
Cursor
.cursor/mcp.json
{
  "mcpServers": {
    "outworx-acme": {
      "url": "https://docs.outworx.io/api/mcp/acme",
      "headers": {
        "Authorization": "Bearer <token>"
      }
    }
  }
}
Cline
VS Code · settings.json
"cline.mcpServers": {
  "outworx-acme": {
    "url": "https://docs.outworx.io/api/mcp/acme",
    "headers": {
      "Authorization": "Bearer <token>"
    }
  }
}
Continue
.continue/config.json
{
  "mcpServers": [
    {
      "name": "outworx-acme",
      "transport": "http",
      "url": "https://docs.outworx.io/api/mcp/acme"
    }
  ]
}
Web-based clients connect over OAuth. Claude.ai web, ChatGPT, and any other MCP client that doesn't take a config file go through the standard OAuth 2.1 flow — pick a project on the consent screen and the assistant pulls a scoped token automatically. Per-tool permissions still apply.

The 12 tools

Discovery, codegen, live ops, Q&A, audit.

Discovery

  • list_endpoints

    Browse the spec.

  • search_endpoints

    Natural-language search.

  • get_endpoint

    Schema + parameters.

  • list_versions

    Multi-version aware.

Code generation

  • generate_types

    Whole spec or one endpoint.

  • generate_client

    TS or Python SDK file.

  • generate_example

    cURL, Python, Go, etc.

Live ops

  • execute_endpoint

    Real call, user's creds.

  • get_response_schema

    Inferred from samples.

Q&A

  • ask_docs

    RAG-grounded chat as a tool.

Audit

  • list_recent_calls

    Last 50 invocations.

  • get_activity

    Per-tool / per-day stats.

Self-healing TypeScript

Types that improve as the API runs.

When the spec under-documents a response, MCP calls infer the shape from real traffic. Every sample merges into the stored schema, so generated TypeScript stays accurate as your API evolves — without you re-authoring the spec.

12 tools out of the box

Standard MCP, fully scoped to one project. Per-tool permission toggles for destructive ops.

Sparse-spec recovery

An OpenAPI file with missing response schemas slowly fills in. Every live call captures, samples, and merges.

Drift detection on every call

If the API returns a new field or changes a type, the next call surfaces it, widens the union, and invalidates the cached types.

Calls go through user's credentials

Outworx never sees the credential. Per-tool permission toggles + per-IP rate limits cap the blast radius.

Live call → schema merge
Spec says
{
  "id": string,
  "email": string
}
Real response
{
  "id": "u_8f3a",
  "email": "ada@acme.io",
  "created_at": "2026-05-07T10:32:00Z",
  "preferences": { "theme": "dark" }
}
Merged schema
{
  "id": string,
  "email": string,
  /** sampled */
  "created_at": string,
  /** sampled */
  "preferences": { "theme": string }
}
Next call to generate_types emits the wider type. The cached version invalidates if the next sample widens the schema again.

Pairs well with

AI Chat

Same retrieval engine, exposed as a public docs drawer. Visitors get answers; engineers get the MCP tools.

SDK Generator

MCP's generate_client tool emits the same drop-in TS / Python SDK you'd download from the dashboard.

Mock Server

When the model wants to iterate without touching prod, point execute_endpoint at the mock URL instead.

Make your API native to AI assistants.

Pro plan unlocks an MCP server per project. Connect it to Claude, Cursor, Cline, or Continue in 30 seconds.

Start freeSee pricing