# Use an MCP tool with a MandAPI-backed coding session

Keep model inference in MandAPI while Claude Code connects to an MCP server locally.

## Keep the two connections clear

MandAPI supplies model inference. Your coding client connects to MCP servers and executes their tools; this guide does not require a MandAPI-hosted MCP endpoint. The model still needs tool-calling support in the API format used by your client.

Start with the Claude Code MandAPI connection already working: `ANTHROPIC_BASE_URL=https://api.mandapi.com`, a private `ANTHROPIC_AUTH_TOKEN`, and an account-visible Anthropic Messages model. Other OpenAI-compatible clients normally use `https://api.mandapi.com/v1` for inference. Those model URLs are not MCP server addresses.

## Add one fetch tool

This example uses the reference Fetch MCP server to read a public page. Install `uv` from its official instructions if it is not already available; `uvx` launches the server package. From your project directory, register a local-scoped server:

```bash
claude mcp add --transport stdio --scope local web-fetch -- uvx mcp-server-fetch
claude mcp get web-fetch
```

The server registration launches the fetch process on your computer. It does not need your MandAPI key as a command argument.

## Run a visible tool check

Start Claude Code with the MandAPI model selected and run `/mcp` to inspect the server. Ask: “Use web-fetch to read https://mandapi.com/en/ and summarize its main product in two sentences. Show the source URL.” Review the normal tool confirmation when shown.

Check that a fetch result appears and the answer uses that result. Then inspect MandAPI usage for the model request. Server startup, successful fetch and successful model inference are three separate checks.

## Fix or remove the server

If the MCP server fails to start, check that the launched process can find `uvx`. On Windows, follow the Fetch server's UTF-8 guidance for encoding-related timeouts. If fetching works but the model cannot request tools, check the selected model's tool support and the inference connection.

Remove only this example with `claude mcp remove web-fetch --scope local`. Keep other configured MCP servers intact.

## References

- [https://code.claude.com/docs/en/mcp](https://code.claude.com/docs/en/mcp)
- [https://github.com/modelcontextprotocol/servers/tree/main/src/fetch](https://github.com/modelcontextprotocol/servers/tree/main/src/fetch)
- [https://modelcontextprotocol.io/docs/learn/architecture](https://modelcontextprotocol.io/docs/learn/architecture)
- [https://openrouter.ai/docs/cookbook/coding-agents/mcp-servers](https://openrouter.ai/docs/cookbook/coding-agents/mcp-servers)