# veille-mcp

[![CI](https://github.com/Kestrall/veille-mcp/actions/workflows/ci.yml/badge.svg)](https://github.com/Kestrall/veille-mcp/actions/workflows/ci.yml)

An [MCP](https://modelcontextprotocol.io) server that turns a curated set of RSS feeds into a **Data, AI and cloud tech watch** any MCP client can use (Claude Code, Claude Desktop, Cursor…), with digests written by a **local LLM served by Ollama**: no article leaves the network to be summarised.

It runs in my home lab, in a Proxmox LXC container, next to the Ollama server.

```
Claude Code ──MCP / Streamable HTTP──▶ veille-mcp (LXC) ──▶ 16 RSS feeds
                                            │
                                            └──▶ Ollama qwen3:8b (LXC, local LLM)
```

## What it exposes

| Kind | Name | Purpose |
|------|------|---------|
| Tool | `latest_news` | Newest articles from the last *n* hours, filtered by category (`ia`, `data`, `cloud`) or source |
| Tool | `search_news` | Keyword search over titles and summaries, title matches ranked first |
| Tool | `digest` | Summary written by the local LLM: key facts with numbered citations, trends, impact for a European cloud provider |
| Tool | `list_sources` | Feeds being watched |
| Resource | `veille://sources` | Feed configuration as JSON |
| Prompt | `competitive_watch` | Ready-made competitive brief on a vendor (AWS, Scaleway, Google Cloud…) |

Tools return **typed, structured output** (Pydantic models) and are annotated read-only. Feeds are fetched concurrently and cached for 15 minutes; a failing feed is reported in `unavailable_sources` instead of breaking the call.

Default feeds cover AI labs and research (Hugging Face, OpenAI, Google AI, Simon Willison, The Decoder, MIT Technology Review…), data engineering (InfoQ, Databricks) and cloud vendors (OVHcloud, Scaleway, AWS, Google Cloud, Azure). Point `VEILLE_SOURCES` at your own JSON file to change them.

## Use it

```sh
pip install -e .
veille-mcp                                   # stdio, for a local client
OLLAMA_URL=http://localhost:11434 veille-mcp --transport streamable-http --host 0.0.0.0
```

Add it to Claude Code:

```sh
claude mcp add --transport http veille http://<host>:8000/mcp
```

Then just ask: *"What happened in AI this week? Give me a digest focused on European cloud providers."*

| Variable | Default | |
|----------|---------|---|
| `OLLAMA_URL` | *(unset)* | Ollama endpoint; `digest` is disabled without it |
| `OLLAMA_MODEL` | `qwen3:8b` | Any model pulled in Ollama |
| `VEILLE_ALLOWED_HOSTS` | *(unset)* | Comma-separated `host:port` list; enables DNS-rebinding protection |
| `VEILLE_SOURCES` | bundled list | Path to a JSON feed list |
| `VEILLE_CACHE_TTL` | `900` | Feed cache, in seconds |

`digest` runs on CPU in my lab and takes a few minutes; raise the client timeout if needed (`MCP_TOOL_TIMEOUT=600000` for Claude Code).

## Deploy (Proxmox LXC)

```sh
tar --exclude=.venv --exclude=.git -czf - . | ssh root@<ct> 'mkdir -p /opt/veille-mcp && tar -xzf - -C /opt/veille-mcp && /opt/veille-mcp/deploy/install.sh'
```

`deploy/install.sh` creates a dedicated system user, a virtualenv and a hardened systemd unit (`ProtectSystem=strict`, `NoNewPrivileges`). Settings live in `/etc/veille-mcp.env`.

## Develop

```sh
pip install -e ".[dev]"
ruff check . && pytest --cov
```

Tests run the server in-process through the MCP client, with mocked HTTP for feeds and Ollama. CI runs lint and tests on Python 3.12 and 3.13.

## License

MIT