With Claude Code and other agents

arvo-mcp-server is Arvo as an MCP server over stdio. An agent gets the whole research surface — write a rule, run a study, read the findings — and cannot fetch data, reach a broker, place an order or share anything.

claude mcp add arvo -- ~/.local/arvo/arvo-mcp-server

Then, in Claude Code:

Run a study of sma_cross on MSFT.YF over the last ten years and tell me whether the verdict survives conservative costs.

Say who is running

claude mcp add arvo -- ~/.local/arvo/arvo-mcp-server --agent claude

--agent <name> matters more than it looks. Findings are saved under that author, and deflated against everything that author has ever run. An agent that runs study after study until one comes out Supported has performed a search, and the winner of twenty studies is the best of a hundred and eighty draws even if each study was honestly deflated against its own nine trials. Naming the author is what lets the engine hold the whole loop to one bar.

Every run is also appended to agent-audit.jsonl in the project, so an unattended agent leaves a trail you can read afterwards.

Why it cannot trade

This is structural, not a prompt.

flowchart LR
    A["Agent<br/><small>Claude Code, Codex, …</small>"]
    M["arvo-mcp-server"]
    R["Research service<br/><small>studies · findings · library reads</small>"]
    C["Control service<br/><small>sessions · halt · reconcile</small>"]
    V["Venues and vendors"]

    A -->|stdio, MCP| M
    M -->|research token<br/>engine.json| R
    M -.->|"no token"| C
    C --> V

    linkStyle 2 stroke:#d1242f,stroke-dasharray:4 4
    style C fill:#fff1e5,stroke:#d1242f

Three things hold the line:

1. The server holds the research token only. control.json — the token that can start a session or halt one — is never given to it.
2. A test enumerates every offered tool by name and fails the build if one appears containing fetch, order, trade, share, import, key or sign. Adding a tool that could reach money breaks CI.
3. The research service has no such call to make. A parallel test on the service itself asserts the same thing from the other side.

It reads the library; it does not fill it

query_market_data returns bars already on disk. Nothing an agent can call contacts a vendor, so an agent cannot run up an API bill or change what a later study will read.

The tools

Tool	What it does
list_strategies	the rules Arvo ships, and what the project and its extensions offer
list_rules / write_rule	rules as data; write one, refused with the reason when it cannot run
list_rulesets / write_ruleset	parameter grids over a rule
translate_pine	read a Pine v5 strategy as a rule, refusing what it cannot say by name
list_instruments	what the library holds
list_findings / open_finding	research memory, and one finding in full
run_study	one rule, one instrument, a search over its parameters, a verdict
run_panel	one rule across a universe — the same configuration for every member
run_walk_forward	re-select on a rolling schedule instead of splitting once
query_market_data	bars for an instrument over a window, at most 2000 from the end
inspect_regime	the regime each bar closed in, labelled after the fact
rank_findings	the leaderboard, in the one order
compare_experiments	findings read against each other — then deflated again, because keeping the best of six is a search of size six
read_review	what every session and every person did on a day, as one report

How it finds the engine

It reads engine.json in the app data directory. If no engine is running it starts one — from ARVO_ENGINE, or arvo-engine beside itself — and that engine keeps running after the agent's session ends, so a study you asked for at midnight is still there in the morning.

To point an agent at a specific project rather than the app data directory, give it the folder:

claude mcp add arvo -- ~/.local/arvo/arvo-mcp-server ~/arvo --agent claude

Other clients

Any MCP client works the same way — the server names no vendor and assumes no particular agent.

Claude CodeCodex / generic MCP

claude mcp add arvo -- ~/.local/arvo/arvo-mcp-server --agent claude

{
  "mcpServers": {
    "arvo": {
      "command": "/home/you/.local/arvo/arvo-mcp-server",
      "args": ["/home/you/arvo", "--agent", "codex"]
    }
  }
}

The window's Agent panel is this same server attached over the Agent Client Protocol, so an agent sees exactly the same tools inside Arvo and outside it.

What an agent is for, and what it is not

The agent's job is to write and test the rules that trade, and to operate them — not to place trades of its own. Starting a session is a person's decision, made with the control token, and promotion to real money needs five paper days on the record regardless of who asks.

The arvo-skills plugin teaches the agent how to use these tools the way Arvo means them — verdict before number, every run a search.

Where this is going →