What does this MCP prompt chain do?

It connects Claude to a Model Context Protocol (MCP) server through promptfoo, automatically executing any tool calls Claude emits and feeding the results back into the conversation until Claude produces a final reply.

How do I set up and run this example?

Set your ANTHROPIC_API_KEY environment variable, then run `npx promptfoo@latest init --example anthropic/mcp`, navigate to the directory, and execute `npx promptfoo@latest eval`.

Can I use a different MCP server instead of the default deepwiki server?

Yes, you can replace the `mcp` block in the config to point to a local stdio server or a remote SSE/streamable HTTP server instead of the bundled deepwiki server.

What happens if a tool call fails?

Tool errors are returned as `tool_result` blocks with `is_error: true`, allowing Claude to recover from or report the failure.

What is the default limit for tool calls per request?

The default `max_tool_calls` limit is 8, which bounds the cost of multi-round tool execution.

Prompt Chain

Integrate LLMs with External Tools via MCP

Name: Mcp
Availability: OnlineOnly
Author: Promptfoo

Promptfoo example wiring Claude to an MCP server via the Anthropic Messages provider, auto-executing tool calls.

Copy chain

Works with anthropic

Promptfoo

Own this? Claim it

Spark score

out of 100

Updated last month

Version code-scan-action-0.1

Models

claude

Add to Favorites

Why it matters

Connect large language models like Claude to external tools and services using the Model Context Protocol. This asset enables LLMs to leverage tool outputs and integrate them back into conversations for more dynamic interactions.

Outcomes

What it gets done

Wire Anthropic's Claude to an MCP server.

Expose MCP server tools to the LLM.

Execute tool_use blocks emitted by the model.

Feed tool results back into the conversation.

Install

Add it to your toolbox

Run in your project directory:

curl -fsSL https://spark.entire.vc/get/pfoo-mcp | bash

Steps

Steps in the chain

Initialize MCP example project

Set up API key and run evaluation

Configure MCP server (optional)

Overview

Mcp

A promptfoo example wiring Claude to an MCP server through the Anthropic Messages provider, with automatic multi-round tool_use/tool_result execution. Use when you want promptfoo to auto-execute MCP tool calls during a Claude eval instead of manually scripting the tool-call loop.

What it does

This example wires Claude up to a Model Context Protocol (MCP) server through the Anthropic Messages provider. Promptfoo exposes the MCP server's tools to Claude, executes any tool_use blocks the model emits, and feeds the tool_result back into the conversation until Claude produces a final reply. The bundled config uses the public deepwiki MCP server so the example works out of the box without installing anything.

When to use - and when NOT to

Use this example when you want promptfoo to automatically execute MCP tool calls that Claude emits during an eval, rather than manually scripting the tool-call loop yourself. It is not specific to a single MCP server - the bundled deepwiki server is just a zero-setup default; swap the mcp block to point at a local stdio server or a remote SSE/streamable HTTP server instead.

Inputs and outputs

Set ANTHROPIC_API_KEY, then scaffold and run:

npx promptfoo@latest init --example anthropic/mcp
cd anthropic/mcp
export ANTHROPIC_API_KEY=sk-ant-...
npx promptfoo@latest eval

To swap in a custom server, replace the mcp block, e.g. a local stdio server: mcp: { enabled: true, server: { command: npx, args: ['-y', '@modelcontextprotocol/server-filesystem', '/tmp/workspace'] } }, or a remote server under a servers list with name/url.

Integrations

Demonstrates tool discovery (MCP-exposed tools are forwarded to Claude alongside any inline tools defined in config), multi-round execution (when Claude returns a tool_use block matching an MCP tool, promptfoo invokes it with the model's arguments and appends a tool_result on the next turn, looping until Claude returns text, bounded by max_tool_calls, default 8), error pass-through (tool errors return as tool_result blocks with is_error: true so Claude can recover or report the failure), and mixed tool handling (non-MCP tools, including built-ins like web_search, fall through to the existing tool-use output without being auto-executed). Promptfoo's disk response cache is automatically skipped when mcp.enabled is true, since tool results can vary between runs; max_tool_calls bounds per-request cost instead. The same pattern is shown elsewhere via the OpenAI Responses API (examples/openai-mcp), and an MCP server can be tested directly without an LLM provider via examples/simple-mcp.

Who it's for

Teams building Claude-based agents that call MCP tools who want promptfoo to handle the full tool-call execution loop automatically during evals, rather than hand-writing the multi-round tool_use/tool_result exchange themselves.

Source README

anthropic/mcp (Anthropic Messages + MCP tools)

This example wires Claude up to a Model Context Protocol server through the Anthropic Messages provider. Promptfoo exposes the MCP server's tools to Claude, executes any tool_use blocks the model emits, and feeds the tool_result back into the conversation until Claude produces a final reply.

npx promptfoo@latest init --example anthropic/mcp
cd anthropic/mcp

Setup

export ANTHROPIC_API_KEY=sk-ant-...
npx promptfoo@latest eval

The bundled config uses the public deepwiki MCP server so the example works out of the box without installing anything. To swap in your own server, replace the mcp block:

mcp:
  enabled: true
  # Local stdio server
  server:
    command: npx
    args: ['-y', '@modelcontextprotocol/server-filesystem', '/tmp/workspace']
  # …or remote SSE / streamable HTTP
  # servers:
  #   - name: my-server
  #     url: https://example.com/mcp

What it demonstrates

Tool discovery: Tools exposed by the MCP server are forwarded to Claude alongside any inline tools you define.
Multi-round execution: When Claude returns a tool_use block matching an MCP tool, promptfoo invokes the tool with the model's arguments and appends a tool_result on the next user turn. The loop continues until Claude returns text - bounded by max_tool_calls (default 8).
Error pass-through: Tool errors come back as tool_result blocks with is_error: true so Claude can recover or report the failure.
Mixed tool handling: Non-MCP tools (regular function-calling tools, or built-ins like web_search) fall through to the existing tool-use output without being auto-executed.

Caching

Promptfoo's disk response cache is skipped automatically when mcp.enabled is true, because tool results can vary between runs. Use max_tool_calls to bound the per-request cost.

Common questions

Discussion