Muse Glimmer + MCP + Langfuse (local)

A minimal Pydantic AI agent that:

• runs muse-glimmer (Meta's 30B agentic model) through a local Ollama instance,
• has MCP enabled: a local FastMCP server (mcp_server.py) is attached as a stdio subprocess, exposing get_weather, get_current_time, and list_ollama_models as agent tools,
• sends full traces (model requests, tool calls, outputs) to a local Langfuse instance for observability.

Prerequisites

• Ollama running with the model pulled:

  ollama pull muse-glimmer
  

• A local Langfuse instance (e.g. via docker compose from the Langfuse repo) reachable at http://localhost:3000.
• uv (or use python3 -m venv + pip).

Setup

cp .env.example .env   # then fill in your Langfuse keys
uv sync                # installs pydantic-ai, fastmcp, langfuse, ...

.env:

LANGFUSE_PUBLIC_KEY=pk-lf-...
LANGFUSE_SECRET_KEY=sk-lf-...
LANGFUSE_BASE_URL=http://localhost:3000
LANGFUSE_HOST=http://localhost:3000

OLLAMA_BASE_URL=http://localhost:11434/v1
OLLAMA_MODEL=muse-glimmer

Run

uv run python agent.py

The agent answers three demo prompts, calling MCP tools as needed:

🧑  User: What is the weather in Paris today?
🤖  Agent: Weather in Paris: clear skies, 22°C, humidity 51%.

🧑  User: What time is it in Tokyo right now?
🤖  Agent: The current time in Asia/Tokyo is 2026-08-25 21:04:33 JST.

🧑  User: Which Ollama models are available locally?
🤖  Agent: NAME  ID  SIZE  MODIFIED ...

Observability (Langfuse)

Open http://localhost:3000 → Traces. Each run produces traces named muse-glimmer-mcp-agent, with spans for model requests (input/output tokens) and each MCP tool call — inspect inputs, outputs, latencies, and costs.

Expose the agent as an MCP server

agent_mcp_server.py flips the architecture around: the agent itself becomes an MCP server exposing one tool — ask_agent(prompt) — to any MCP client (Claude Desktop, VS Code, Cursor, another Pydantic AI agent, ...). Each call runs muse-glimmer via Ollama, still has the inner MCP tools, and is traced to Langfuse.

Run the server

uv run python agent_mcp_server.py

Client configuration

Point any MCP client at this server using your uv-managed venv's Python (uv run --project <repo> python also works). Examples:

Claude Desktop — claude_desktop_config.json:

{
  "mcpServers": {
    "muse-glimmer-agent": {
      "command": "/home/d3lee/.local/bin/uv",
      "args": ["run", "--project", "/home/d3lee/my-repos/pydantic-ai-mcp-server-sample", "python", "agent_mcp_server.py"]
    }
  }
}

VS Code — .vscode/mcp.json:

{
  "servers": {
    "muse-glimmer-agent": {
      "type": "stdio",
      "command": "uv",
      "args": ["run", "--project", "/home/d3lee/my-repos/pydantic-ai-mcp-server-sample", "python", "agent_mcp_server.py"]
    }
  }
}

Cursor — .cursor/mcp.json:

{
  "mcpServers": {
    "muse-glimmer-agent": {
      "command": "uv",
      "args": ["run", "--project", "/home/d3lee/my-repos/pydantic-ai-mcp-server-sample", "python", "agent_mcp_server.py"]
    }
  }
}

Layout

File	Purpose
agent.py	Pydantic AI agent: Ollama model + MCP capability + Langfuse
mcp_server.py	FastMCP server (stdio) exposing the local tools
agent_mcp_server.py	Exposes the agent itself as an MCP server (ask_agent tool)
.env	Langfuse + Ollama configuration

pydantic-ai-mcp-server-sample