> ## Documentation Index
> Fetch the complete documentation index at: https://docs.anpord.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Mock MCP servers

> Test agent tool use without external credentials or data

Define typed tools and resources with Standard Schema, then attach the server to
an eval:

```ts theme={null}
import { defineEval } from "anpord";
import { server, tool } from "anpord/mcp";
import { z } from "zod";

const users = server({
  name: "users",
  version: "1.0.0",
  tools: [
    tool({
      name: "users_get",
      description: "Get a user by ID",
      inputSchema: z.object({ id: z.string() }),
      outputSchema: z.object({ id: z.string(), name: z.string() }),
      handler: ({ id }) => ({ id, name: "Ada" }),
    }),
  ],
});

export default defineEval({
  name: "uses-users-api",
  mcp: [users],
  // cases, prompt, source, tasks, and trials
});
```

The same definition works with Codex, Claude, OpenCode, Pi, Gemini, Qwen,
Cursor, and FX. Anpord bundles a fresh stdio process into each task and writes
the harness's native configuration. Pi's MCP adapter is installed automatically.
The custom `command` harness owns its process and MCP integration, so it does
not receive an implicit configuration.

Handlers use plain TypeScript and may return a value or promise. Inputs and
outputs are validated against their schemas. No external organization IDs,
service credentials, or test data are required.

Inspect exact calls in a validator:

```ts theme={null}
import type { Validator } from "anpord";

export const calledUser: Validator = async ({ mcp }) => ({
  message: "the agent called users_get",
  passed: (await mcp.calls("users")).some(
    (call) => call.kind === "tool" && call.name === "users_get"
  ),
});
```