web-mcp

Semantic search over RSS/Atom feeds for AI agents — a single-file MCP server. No API key, no cloud, just the feeds you choose.

Point it at a handful of RSS/Atom feeds and give an agent tools to search them by meaning (TF-IDF + recency boost), see what's trending, generate a digest, and monitor feed health — all running locally against feeds you control.

Features

• Add/remove/list feeds — RSS 2.0 and Atom

• Ranked search — TF-IDF cosine similarity with a recency boost, pure stdlib (no embeddings, no network calls beyond fetching the feeds themselves)

• Trending topics — most frequent terms across recent articles

• Source health — flags feeds that are erroring, silent, or stale

• Markdown digests — grouped-by-feed summary of recent articles

• Background refresh — feeds are refreshed on a timer automatically

Install

Requires Python 3.10+.

On the package name: the project is web-mcp, but the PyPI distribution is feed-mcp — PyPI rejects web-mcp as too similar to an unrelated existing webmcp package. So you pip install feed-mcp, and it installs both a web-mcp and a feed-mcp command; they're the same program, use whichever you prefer.

Option A — Claude Code plugin (recommended)

Installs the MCP server and the bundled skills and agents in one step:

pip install feed-mcp

Then, inside Claude Code:

/plugin marketplace add ManiaSacha/web-mcp
/plugin install web-mcp@maniasacha-web-mcp

The plugin registers the web MCP server automatically and prompts you for a starting feed list. If the install summary says Run /reload-plugins to activate., run that.

The pip install step is required: the plugin's MCP server invokes the web-mcp command, which pip puts on your PATH.

Option B — MCP server only

pip install feed-mcp
claude mcp add web -- web-mcp --feeds https://hnrss.org/frontpage

For Claude Desktop, add this to claude_desktop_config.json:

{
  "mcpServers": {
    "web": {
      "command": "web-mcp",
      "args": ["--feeds", "https://hnrss.org/frontpage,https://github.blog/feed/"]
    }
  }
}

Option C — from source

git clone https://github.com/ManiaSacha/web-mcp.git
cd web-mcp
./install.sh          # Windows: .\install.ps1

The install scripts check your Python version, install the package, and print the exact next command for your setup.

Quick start

Run the server directly to check it works:

web-mcp --feeds https://hnrss.org/frontpage,https://github.blog/feed/

It speaks MCP over stdio, so it will sit silently waiting for a client — that's correct behavior, not a hang. Press Ctrl+C to exit.

Once connected, ask your agent things like:

• "What's trending across my feeds today?"

• "Search my feeds for anything about Postgres performance."

• "Give me a digest of the last 24 hours."

• "Are any of my feeds broken?"

What's included

Tools

Security notes

add_feed(url) makes the server fetch a URL that may have reached your agent from a web page rather than from you, so that path is treated as untrusted:

• Scheme allowlist — only http:// and https://. file://, ftp://, gopher://, and data: are refused.

• Public addresses only — a URL is rejected unless every address it resolves to is publicly routable. Loopback, private, link-local (including the 169.254.169.254 cloud metadata endpoint), multicast, and reserved addresses are blocked, as are their IPv4-mapped IPv6 forms.

• Connect-time IP pinning — the address that was validated is the one actually dialed, so DNS cannot answer differently between the check and the connection (rebinding). TLS still validates certificates against the hostname.

• Per-hop redirect validation — redirects are followed manually and re-validated at every hop, so a public URL cannot bounce the fetch to an internal host. Capped at 5 hops.

• Resource bounds — 5 MB per feed body, an explicit socket timeout, and a clamp on every tool's limit.

Two things this does not do: it has no authentication (it's meant to run locally over stdio — don't expose it on a network), and it does not vet feed content. Indexed article text is written by whoever controls the feed, so treat it as data, never as instructions.

Full detail, including the limits that remain, is in docs/SECURITY-MODEL.md. To report a vulnerability, see SECURITY.md.

Development

pip install -e ".[dev]"
pytest              # 53 tests, no network access
ruff check .        # lint
python scripts/validate_plugin.py   # manifests, frontmatter, version sync

CI runs all three on Python 3.10–3.13 (Linux, plus macOS and Windows on 3.12), and additionally installs the built wheel to confirm the console scripts work.

See CONTRIBUTING.md for the workflow and CHANGELOG.md for release history.

License

MIT