# MCP Status Panel — Design

Date: 2026-08-03
Status: Approved

## Problem

`app/mcp/` already exposes the compliance knowledge base over MCP
(`search_regulations`, Streamable HTTP at `/mcp/`, JWT-guarded, Host
allow-listed). It is completely invisible from the product: an operator
looking at the System Status page cannot tell whether the MCP endpoint is
enabled, what URL a client should point at, which tools are advertised, or
whether anything has ever called it.

This design adds that visibility, and only that.

## Goals

- Show MCP endpoint configuration (public URL, auth on/off, Host allow-list).
- Show the advertised tool list, read from the live MCP registry rather than
  hard-coded.
- Show per-tool call counters: total calls, errors, average duration, last
  call time.
- Give the operator a one-click "copy client config" JSON they can paste into
  Claude Desktop / Cursor.

## Non-Goals

- No persistence. Counters are in-memory and reset on restart. Token
  consumption caused by MCP calls is already persisted by the existing
  `ModelUsageTracker` (MCP calls route through `AgentConversationService.ask()`
  like every other caller), so nothing billable is lost.
- No historical trends or time-series charts.
- No active-client / session list. That would require hooking the MCP SDK's
  internal `StreamableHTTPSessionManager`, which is private API and breaks on
  SDK upgrades.
- No per-client attribution.
- No new frontend test framework — the project has none, and this change does
  not justify introducing one.

## Architecture

Module boundary is unchanged: everything new on the backend lands inside the
existing `app/mcp/` module, plus one thin HTTP adapter route.

```
frontend/src/pages/Status/StatusPage.tsx
        │  GET /api/v1/status/mcp
        ▼
backend/app/api/routes/status.py        ← HTTP adapter only
        │  get_mcp_status(public_url)
        ▼
backend/app/mcp/server.py               ← assembles the status payload
        ├── settings (endpoint / auth / allowed hosts)
        ├── mcp.list_tools()            ← live tool registry
        └── app/mcp/stats.py            ← in-memory counters
```

The status route must not reach into the MCP server's internals. It passes in
the resolved public URL (the one thing only the HTTP layer knows) and receives
a finished dict. This keeps the MCP protocol details in one module.

## Components

### `app/mcp/stats.py` (new)

Mirrors `app/shared/model_usage_tracker.py` in shape and in defensive posture.

- `MCPToolStats` dataclass: `calls`, `errors`, `last_called_at`,
  `total_duration_ms`; `avg_duration_ms` computed as a property.
- `MCPStatsTracker`: one `threading.Lock` guarding a
  `dict[str, MCPToolStats]`. `record(tool, duration_ms, success)` and
  `snapshot()`.
- `get_mcp_stats_tracker()`, `@lru_cache` singleton.

The lock is not optional. The mcp SDK (2.0.0) dispatches synchronous tool
functions via `anyio.to_thread.run_sync`, so `search_regulations` genuinely
runs on multiple worker threads concurrently — unlike the async REST routes,
which serialize on the event loop.

`record()` swallows and logs its own exceptions, matching `ModelUsageTracker`:
a defect in observability code must never fail a real MCP tool call.

### `app/mcp/server.py` (modified)

- `search_regulations` gets a `try/except/finally` wrapper that measures
  elapsed time with `time.perf_counter()` and records success or failure. The
  exception is re-raised after recording — the MCP SDK still needs to turn it
  into a protocol-level error.
- New `async def get_mcp_status(public_url: str) -> dict` merges three
  sources: settings, `await mcp.list_tools()`, and the stats snapshot. Tools
  are matched to their stats by name; a tool that has never been called
  reports zeros.

### `app/api/routes/status.py` (modified)

`GET /status/mcp` resolves the public URL, then delegates:

```python
public_url = settings.mcp_public_url or f"{str(request.base_url).rstrip('/')}/mcp/"
return await get_mcp_status(public_url)
```

### `app/config/settings.py` + `.env.example` (modified)

New optional `mcp_public_url: str = ""`.

This override is required, not cosmetic. The Vite dev proxy sets
`changeOrigin: true` (`frontend/vite.config.ts`), which rewrites the `Host`
header to the proxy target, so `request.base_url` on the backend reads
`http://127.0.0.1:8000/` in development regardless of how the operator
actually reached the page. Deployments behind a reverse proxy that does not
forward the original Host have the same problem. When unset, the derived
value is correct for the common same-origin case.

### Frontend

- `api/index.ts`: `MCPToolEntry` and `MCPStatusResponse` types, alongside the
  existing `ModelUsageEntry` / `SystemHealth` types.
- `api/status.ts`: `getMCPStatus()`, using the typed `fetchAPI` client.
- `StatusPage.tsx`: new "MCP Server" card in the left column, directly below
  the existing "AI Models" card. It joins the existing
  `Promise.allSettled([...])` batch, so it refreshes with the page's existing
  Refresh button and needs no independent polling. `handleExport()` includes
  the MCP payload.
  - Header row: title + "copy client config" button.
  - Endpoint row: `StatusIcon` + monospace URL + auth badge + Host allow-list.
  - One row per tool: name, calls, errors, average duration, last call time.
  - Reuses the existing `card`, `card-header`, `service-row` classes and the
    `StatusIcon` component. No new CSS.
- `locales/zh.ts` / `locales/en.ts`: new keys under the existing `status`
  section.

### Copy client config

Produces the Streamable HTTP form both Claude Desktop and Cursor accept:

```json
{
  "mcpServers": {
    "ai-regulations": {
      "url": "http://6.86.80.9:8000/mcp/",
      "headers": { "Authorization": "Bearer <token>" }
    }
  }
}
```

The real JWT from `localStorage` is embedded, because a config with a
placeholder does not work when pasted and defeats the button's purpose. This
is the operator's own token, already present in their own browser; the button
moves it from one local store to another local store on the same machine. The
`headers` key is omitted entirely when `auth_required` is false.

## Data Flow

1. StatusPage mounts (or Refresh is pressed) → `getMCPStatus()` in the
   existing parallel batch.
2. Route resolves the public URL and calls `get_mcp_status()`.
3. `get_mcp_status()` reads settings, awaits `mcp.list_tools()`, snapshots
   stats, joins tools to stats by name.
4. Card renders. Counters advance only when a real MCP client calls a tool.

## Error Handling

- `GET /status/mcp` fails or times out → `Promise.allSettled` leaves the state
  `null` → card renders a muted "unavailable" body. This is the same pattern
  the existing model-usage card already uses; one failing status endpoint must
  never blank the whole page.
- `mcp.list_tools()` reads an in-memory registry populated at import time and
  has no failure mode worth special-casing; an unexpected exception surfaces
  as a 500 on this one endpoint and is contained by the point above.
- `MCPStatsTracker.record()` never raises (logged and swallowed).
- `navigator.clipboard.writeText` rejects on insecure origins and when
  permission is denied. The button reports failure in its own label rather
  than throwing — a silent no-op would leave the operator believing they
  copied something.

## Testing

`backend/tests/mcp/test_mcp_stats.py`:

- Concurrent `record()` from multiple threads yields an exact total (proves
  the lock).
- `avg_duration_ms` is correct across several calls, and is `None` with zero
  calls (no division by zero).
- Successes and failures land in `calls` vs `errors` correctly.
- `record()` on malformed input logs instead of raising.

`backend/tests/mcp/test_mcp_status.py`:

- `get_mcp_status()` returns the advertised tool with zeroed stats before any
  call, and reflects recorded stats after.
- `auth_required` follows `settings.auth_enabled`.
- The passed-in `public_url` appears unmodified in the payload.

Frontend: no new tests; verified via `npm --prefix frontend run lint` and
`npm --prefix frontend run build`.