# LocalPulse Edge

LocalPulse Edge is a privacy-first local MCP service for turning user-authorized CSV, JSON, or XLSX comment exports into risk signals, prioritized comments, evidence-backed actions, and local reports.

This public source release intentionally contains no trained model weights, real comment exports, annotations, runtime logs, competition handoff material, or machine-specific configuration. The included `demo/demo_comments.csv` is synthetic demonstration data.

## Quick start

```powershell
python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt
python scripts\run_demo.py
```

Run the local MCP service over Streamable HTTP:

```powershell
python scripts\run_mcp.py --transport streamable-http --host 127.0.0.1 --port 8890
```

The service endpoint is `http://127.0.0.1:8890/mcp`. It listens only on the local machine in the example above.

## Trae: one-click MCP import

1. Open PowerShell in the project directory and run:

   ```powershell
   powershell -ExecutionPolicy Bypass -File scripts\install_trae_mcp.ps1 -ProjectRoot .
   ```

2. The script starts and verifies the local MCP service, writes a JSON configuration under `artifacts/`, and opens Trae's import link.
3. In Trae, confirm the import and enable the project Skill at `.trae/skills/localpulse/SKILL.md`.

If the link cannot be opened, paste the contents of `.trae/localpulse.http.mcp.json` into Trae's MCP configuration UI. This release uses a standard HTTP MCP server; it does not require NPX or UVX packaging.

## Models and execution evidence

Without local model assets, the service remains usable for schema validation, cleaning, provenance, and deterministic fallback behavior. To enable hardware-accelerated inference, place independently obtained compatible assets outside version control and configure their paths with `LOCALPULSE_*` environment variables. Always use each analysis result's runtime evidence to distinguish model execution from fallback.

## Privacy boundary

- Analyze only comment exports the operator is authorized to use.
- URLs are recognized as provenance only; LocalPulse does not log in to platforms, use cookies, or scrape comments.
- Do not expose an unauthenticated MCP endpoint to the public Internet.

## Development checks

```powershell
python -m pytest -q
python scripts\verify_mcp.py --transport streamable-http --port 8890
python scripts\package_skill.py --output artifacts\localpulse-skill-trae.zip
```

## License

MIT. See [LICENSE](LICENSE).