milton-mcp

An MCP server for Milton's analyst agents. Add it to Claude Code, Claude Desktop, Cursor or your own agent, and the agent can ask Milton for a technical, fundamental or macro read, a risk-checked trade idea, or the full desk on a ticker.

The server is a thin layer over Milton's public API. The agents run on Milton's servers; this package only sends your requests and formats the results. You need a Milton account and an API token.

Install

With uv (no separate install step; uvx fetches and runs it):

uvx milton-mcp --version

Or with pip: pip install milton-mcp.

Get an API token

Sign in at withmilton.ai, open Settings → API tokens and create one. It starts with mlt_. Treat it like a password: it can start runs that bill to your balance.

Configure your client

Claude Code

claude mcp add milton -e MILTON_API_KEY=mlt_your_token -- uvx milton-mcp

Claude Desktop, Cursor and other JSON-configured clients

Claude Desktop: claude_desktop_config.json. Cursor: ~/.cursor/mcp.json or .cursor/mcp.json.

{
  "mcpServers": {
    "milton": {
      "command": "uvx",
      "args": ["milton-mcp"],
      "env": { "MILTON_API_KEY": "mlt_your_token" }
    }
  }
}

Your own agent

Any MCP client works. With the OpenAI Agents SDK, for example:

from agents import Agent, Runner
from agents.mcp import MCPServerStdio

async with MCPServerStdio(
    params={"command": "uvx", "args": ["milton-mcp"], "env": {"MILTON_API_KEY": "mlt_your_token"}},
    client_session_timeout_seconds=70,
) as milton:
    agent = Agent(name="Analyst", instructions="...", mcp_servers=[milton])
    result = await Runner.run(agent, "Give me a swing-trade technical read on NVDA using gpt-6.1-sol.")

A tool call can wait up to 50 seconds for a run to finish, so give your client a request timeout of at least 70 seconds.

Settings

Tools

Every agent tool also takes model (any model Milton offers) and an optional effort (defaults to high, or the model's highest level). horizon is intraday, swing, position or long_term.

What comes back

A short digest the model can read (stance, conviction, summary, key levels, risks, how fresh the data is, cost), the full result as structured content, and links to the run in Milton and in the API.

A run that hasn't finished after 50 seconds comes back as a run_id instead; milton_get_run waits up to 50 seconds per call for it. A desk run always returns its run_id at once.

Repeated calls

MCP clients sometimes re-send a call when a connection drops. An identical call (same tool and inputs) within 10 minutes returns the same run and isn't billed again. A call with any different input starts a new run.

Cost

Runs bill the tokens they use at the chosen model's rates (pricing) against your Milton balance; a desk run bills every analyst it runs. Failed runs aren't billed. Pick a smaller model or a lower effort for cheaper, faster runs.

Running it on your own server

--http serves MCP over Streamable HTTP (stateless, JSON responses), so agents on other machines can reach it:

milton-mcp --http --host 0.0.0.0 --port 8000      # endpoint: http://<host>:8000/mcp

In HTTP mode the server holds no API token. Every request must carry the caller's token as Authorization: Bearer mlt_…, which is passed on to Milton; a request without one gets a 401. Put it behind TLS if it's reachable beyond your own network.

Development

uv sync --group dev
uv run pytest
uv run ruff check . && uv run ruff format --check .

The tools' input schemas, the result schema and the fallback model list are generated from Milton's published API description:

uv run python scripts/generate_schemas.py            # from https://withmilton.ai
uv run python scripts/generate_schemas.py --check    # fails if src/milton_mcp/schemas.json is stale

Releases publish to PyPI when a vX.Y.Z tag matching pyproject.toml's version is pushed.

Disclosures

Milton is not a registered investment adviser, broker-dealer or financial planner. Its outputs are AI-generated general information, not personalized investment advice or a recommendation to buy or sell any security. Each result carries Milton's full disclosures.

License

MIT