oas-execute-mcp

An MCP server that puts a safety layer between a language model and a broker order.

A model calls order_submit. Before anything reaches the broker, the request passes six independent checks, any one of which can refuse it — seven across the write paths, since the stop-loss rule governs modifications rather than opens. Every attempt — allowed or refused — is appended to an audit log with the reason. The broker itself sits behind a swappable backend, so the same tool surface runs against an on-disk simulator or a live MetaTrader 5 terminal without the model knowing which.

Built for an agent that, on its own channel, submitted orders against a prop-firm evaluation account without per-trade human approval, where a mis-sized order was not a failed test.

Why it exists

The interesting part of giving an agent a tool that spends money is not the happy path. It is:

• what the tool returns when it half-succeeds — the order filled but the stop did not attach;

• where a timeout lands — the broker never answered, and you do not know whether it filled;

• how a failure is stopped from being reported as a success by the layer above it.

So the design rule here is that the model's own account of what happened is never trusted. Position state is re-read from the broker on every call. The refusals are explicit, few, and sit directly on the execution path rather than beside it in a prompt.

The safety layer

Seven checks, each able to refuse independently, all logged with a machine-readable reason. Six of them stand between order_submit and the broker; check_sl_widen_block governs modifications, so it never sees an open:

The audit append runs after the gates on every path, allowed or refused — it is the record, not a check: a refusal that leaves no trace is treated as a failure of the system.

safety.py groups these into run_entry_safety_gates, run_modify_safety_gates and run_close_safety_gates, so a close is never blocked by a rule that should only govern an open — the failure mode where a safety system traps you in a position it was meant to protect.

Tool surface

Seven tools: order_submit · order_modify · order_close · close_all · position_list · account_info · health_check.

Architecture

    model tool call
          │
          ▼
   server.py  ── MCP stdio server, tool schemas + dispatch
          │
          ▼
   safety.py  ── 7 gates, audit log, kill switch     ◄── refuses here, before the broker
          │
          ▼
   backends/  ── swappable
      ├── stub.py      on-disk simulation, no broker, deterministic
      └── mt5_demo.py  file bridge → MetaTrader 5 (Wine/macOS) via an MQL5 EA
                       localhost HTTP listener, shared-secret token on the firing route

The MQL5 expert advisor (oas_execute_mcp/mql5/OAS_Bridge.mq5) polls a localhost endpoint and executes on the terminal side. backends/base.py is the interface; adding a broker means implementing it, and nothing above the backend changes.

Install

pip install -e ".[dev]"

Point your MCP client at it:

{
  "mcpServers": {
    "oas-execute": {
      "command": "/path/to/.venv/bin/python",
      "args": ["-m", "oas_execute_mcp.server"],
      "cwd": "/path/to/parent-of-this-package"
    }
  }
}

Backend selection and paths are environment variables:

It ships defaulting to stub. You have to choose to point it at a broker.

No credentials live in this repository, and none should. The MT5 backend follows whatever terminal is already logged in; the account record it reads is documentary (login and server name), carries no password, and belongs outside version control.

Tests

PYTHONPATH=. python3 tests/test_smoke.py     # full submit → modify → close cycle + failure modes
for t in tests/test_*.py; do PYTHONPATH=. python3 "$t"; done   # the whole suite
python3 -m pytest tests/ -q                  # same suite, if you would rather type this

Nine test files, all passing on a clean checkout. Each one is a self-contained checker: it runs its assertions at module scope and exits non-zero on failure, so the suite needs nothing but an interpreter. tests/suite_test.py drives the same scripts as subprocesses so pytest is a working entry point too — it is a bridge, not a second suite, and there is no version of the tests that only one runner sees. They cover the submit/modify/close cycle against the stub, stop-loss anchoring and re-derivation after a charged fill, the bridge client-mode handshake, quote and position routes, and the gate that stops an error-shaped payload from being consumed as an empty position list — a real incident, where a timeout returned {"positions": []} and a downstream reader believed it.

Two tests skip rather than pass when their subject is absent. They exercise a consumer of this server, so they need the host project (OAS_HOST_REPO, OAS_RECONCILE_CMD). A check that silently turns green when the thing it checks is missing is worse than no check.

Status and limits

• The stub backend is complete and deterministic. mt5_demo is in production use against a MetaTrader 5 terminal running under Wine on macOS.

• check_rate_limit is implemented but is not exercised by the test bench — the bench simulates broker timeouts and kill-switch fire, not rate-limit responses. Stated because the distinction between "implemented" and "tested" is the whole subject of this repository.

• The MT5 bridge is macOS/Wine-shaped. A native Windows install needs OAS_MT5_COMMON_FILES set and has not been tested by me.

• Single-terminal by design: the listener is a singleton on one port, and a second server process detects the first and forwards to it rather than competing for the bind.

License

MIT — see LICENSE.