MCP tools | Apiweld

Connect Apiweld to Claude Code, Cursor, or any MCP client so your agent can search APIs, add endpoints, check drift, and fix breaking changes.

apiweld mcp runs Apiweld as a Model Context Protocol server. Your coding agent starts it as a subprocess and talks to it over stdin and stdout. It doesn’t open a port.

Every tool runs the same code as a CLI command, so anything the agent does, you can run again from a shell.

Connect your agent

claude mcp add apiweld -- npx apiweld mcp

Add this to .cursor/mcp.json in your project:

.cursor/mcp.jsonjson

{
  "mcpServers": {
    "apiweld": {
      "command": "npx",
      "args": ["apiweld", "mcp"]
    }
  }
}

Any client that supports stdio servers works. Use npx as the command and apiweld mcp as the arguments. If apiweld is installed in the project, apiweld with the argument mcp works too.

Then run npx apiweld init --agents. It adds a short note to AGENTS.md and CLAUDE.md telling the agent to use Apiweld for third-party APIs instead of writing HTTP calls by hand or installing a full SDK. You can edit the note freely. Apiweld won’t add it again once it’s there.

Tools

Tool	Same as	What the agent gets back
`search_apis`	`apiweld search <query>`	Matching APIs, with a one-line description and how recently the spec changed.
`search_operations`	`apiweld search <query> --ops [--api id]`	Matching endpoints, across all APIs or within one.
`describe_operation`	`apiweld show <api> <operation>`	A short TypeScript-style signature of the endpoint (see below).
`add_operations`	`apiweld add <api> <operation>…`	Updates the config and generates the client. Returns the import path and an example call.
`remove_operations`	`apiweld remove <api> <operation>…`	Removes endpoints and regenerates the client.
`check_drift`	`apiweld check [api]`	Upstream changes to the selected endpoints, labelled breaking, risky, or safe.
`get_heal_plan`	`apiweld heal <api>`	For a breaking change: what changed, which call sites break, and hints for the fix.
`verify`	`apiweld verify`	The type-check result, plus your test command’s result if you configured one.

An operation can be an operation id (PostRefunds) or METHOD /path (POST /v1/refunds).

Compact signatures

Full OpenAPI schemas are long and eat context. describe_operation returns the endpoint in the shape of the code the agent is about to write:

// stripe · POST /v1/refunds · auth: bearer (secret key)
createRefund(body: {
  charge?: string;            // ID of the charge to refund
  payment_intent?: string;
  amount?: number;            // minor units; defaults to the full amount
  reason?: "duplicate" | "fraudulent" | "requested_by_customer";
  metadata?: Record<string, string>;
}): Promise<Refund>           // errors: 400, 401, 402, 404

Nested types stay collapsed. The agent can pass a depth to expand them, the same as apiweld show --depth n.

What the server can change

The server only writes to:

the output directory with the generated clients
apiweld.config.ts
apiweld.lock.json
.apiweld/, which holds reports and the runtime drift log

It never edits your application code. When a breaking change needs call sites updated, the agent makes those edits itself, as normal changes you can review.