# AI Legal Watch MCP

MCP server for append-only maintenance of an AI legal developments tracker stored as CSV.

It validates candidate rows, checks for duplicates, preserves historical CSV formatting, and helps rank possible article candidates without directly browsing or judging legal authority.

## What it does

- Reads a tracker CSV and exposes filtered list/digest views
- Dry-runs candidate additions before writing anything
- Appends exactly one approved row when the tracker fingerprint still matches
- Uses a lock to prevent duplicate concurrent appends
- Suggests article candidates with transparent scoring

## What it does not do

- It does not browse the web or verify the legal source for you
- It does not rewrite existing tracker history
- It does not silently merge or normalize old rows beyond tolerated legacy repairs
- It does not require a content archive, though it can fail open when one is configured

## Requirements

- Node.js `20` or newer
- A local MCP client that can launch a stdio server

## Installation

```sh
cd tools/ai-legal-watch-mcp
npm install
npm test
npm run smoke
```

## MCP client setup

Example stdio configuration:

```json
{
  "mcpServers": {
    "ai-legal-watch": {
      "command": "node",
      "args": ["/absolute/path/to/ai-legal-watch-mcp/server.mjs"],
      "env": {
        "AI_LEGAL_WATCH_WORKSPACE": "/absolute/path/to/workspace"
      }
    }
  }
}
```

The tracker defaults to `ai-legal-watch-tracker.csv` in the chosen workspace. Override it with `AI_LEGAL_WATCH_TRACKER` if needed.

If you want candidate suggestion to exclude already-covered rows using a local content archive, set `AI_LEGAL_WATCH_ARCHIVE_ROOT` or `CLEARON_CONTENT_ARCHIVE_ROOT`.

## Safety model

- `propose_tracker_update` never writes
- `apply_tracker_update` appends only one encoded CSV row and requires the exact SHA-256 fingerprint returned by the proposal
- The tracker lock, fingerprint check, and duplicate check are all repeated while the lock is held
- Existing rows are not reserialized, preserving tolerated historical formatting quirks
- New rows must use legal classifications, not publication workflow labels

## Tools

1. `list_watch_items`
2. `find_duplicate_development`
3. `propose_tracker_update`
4. `apply_tracker_update`
5. `generate_watch_digest`
6. `suggest_article_candidates`

## Example workflow

1. Call `find_duplicate_development` with the topic, development date, and official source URL.
2. Call `propose_tracker_update` with the full candidate row.
3. Verify the underlying primary source and legal classification outside this MCP.
4. Call `apply_tracker_update` with the exact `tracker_sha256` and unchanged `candidate`.
5. If the fingerprint is stale, generate a fresh proposal instead of retrying an old payload.

## Development

```sh
npm test
npm run smoke
npm run syntax
```