# Hermes MCP

Hermes MCP is a FastAPI + MCP supervisor for market interpretation, strategy selection, and Trader control.

## What Hermes does

- builds a market narrative from regime data
- scores strategy fit for each account/market concern
- records decision snapshots for audit and dashboarding
- optionally dispatches Trader control actions through a single guarded write path

## Surface

- MCP transport: `/mcp/sse`
- Health: `/health`
- Dashboard: `/dashboard/`
- Primary tool: `report()`

## Hermes to Trader control path

Hermes reads Trader state via strategy snapshots and writes only through Trader's canonical action tool:

- `apply_control_decision(payload)`

Dispatch is gated locally by:

- `HERMES_ALLOW_ACTIONS`

When the gate is false, Hermes still makes and records decisions, but returns a blocked dispatch result instead of changing Trader state.

## Run

```bash
./run.sh 8590
```

## Test

```bash
./tests.sh
```

## Replay decisions

Hermes now stores a full `replay_input` bundle with each new decision, so stored decisions can be replayed against the current decision engine.

Replay recent decisions:

```bash
python3 scripts/replay_decisions.py --limit 20
```

Only show changed outcomes:

```bash
python3 scripts/replay_decisions.py --limit 50 --only-changed
```

JSON output:

```bash
python3 scripts/replay_decisions.py --limit 50 --json
```

Optional filter by concern:

```bash
python3 scripts/replay_decisions.py --concern-id <concern-id>
```

Note: decisions created before this replay capture existed will not have `replay_input` yet.