Asyncroscopy Agent MCP Server - Asyncroscopy Documentation

The agent MCP server exposes the LLM agent swarm as MCP tools. It is how clients drive the LLM.

asyncroscopy/mcp/agent.py holds the swarm and asyncroscopy/mcp/agent_mcp_server.py wraps it in FastMCP.

Start It¶

Start the device stack and the instrument MCP server first:

uv run startup_scripts/run_servers.py --yaml configs/Spectra300.yaml
uv run startup_scripts/run_mcp.py --yaml configs/mcp.yaml

Then start the agent in another terminal or on the agent computer:

uv run startup_scripts/run_llm.py --yaml configs/gemma-llm.yaml

The default endpoint is:

http://127.0.0.1:8002/mcp

To prompt the swarm from a terminal instead of serving it over MCP:

uv run startup_scripts/run_llm.py --yaml configs/gemma-llm.yaml --interactive

If the agent runs on another computer, set agent.mcp_urls to the instrument MCP machine and agent.http_host to the agent machine’s bind address. Use 0.0.0.0 when clients on other machines need to connect.

YAML Contract¶

agent:
  name: Asyncroscopy_Agent_MCP
  transport: streamable-http
  http_host: 0.0.0.0
  http_port: 8002
  model: "gemma4:31b"
  model_provider: "ollama"
  max_steps: 10
  use_init_chat_model: false
  mcp_urls:
    - "http://127.0.0.1:8000/mcp"
  startup_agents:
    - name: "base"
      system_prompt: "You are a helpful assistant."
      tools: ["list_devices"]

mcp_urls is the only link to the instrument stack. use_init_chat_model: true initializes through LangChain’s init_chat_model using model and model_provider instead of a local Ollama model.

Tools¶

Tool	Signature	Purpose
`query_agent`	`(prompt: str, include_transcript: bool) -> str`	Run the swarm and return its response.
`spawn_agent`	`(config: str) -> str`	Add a worker; returns its name, or `""` on failure.
`list_agents`	`() -> list[str]`	Names of the agents in the swarm.
`list_agent_tools`	`() -> str`	JSON list of the inherited MCP tool names.
`complete`	`(request_json: str) -> str`	OpenAI-compatible single step.
`set_max_steps`	`(max_steps: int) -> int`	LangGraph recursion limit per query.

spawn_agent takes name, system_prompt, and optionally tools (glob patterns matched against inherited tool names, e.g. ["*spectrum"]), model, and description.

query_agent returns the whole run by default: supervisor routing, each tool call and its result, and the generated text, followed by FINAL ANSWER:. Pass include_transcript=False to get just the answer.