X书签知识库MCP服务

x-bookmarks-mcp

Local-first, privacy-focused MCP server that turns exported X/Twitter bookmarks into a searchable, enriched knowledge base for Claude Desktop, Claude Code, and other MCP clients.

The bookmark database, notes, tags, and search index stay local. Optional enrichment makes outbound requests to public links contained in the export so their readable text can be indexed.

Why I Built This

X bookmarks are easy to save and difficult to reuse. This project turns an exported bookmark collection into a local knowledge base that can be searched, tagged, annotated, enriched, and exposed through MCP.

The project is intentionally small: SQLite is the source of the local index, FTS5 provides search, Python handles import/enrichment, and the MCP server exposes a bounded tool surface.

Highlights

• Local-first storage — SQLite database and bookmark data stay on your machine.
• No X API dependency — import from JSON exports instead of requiring X API credentials.
• MCP integration — eight tools for search, browsing, notes, tags, stats, and topic synthesis.
• Full-text search — tweet text, exported thread/article text, notes, tags, and enriched linked content.
• Incremental imports — unchanged bookmarks are skipped; changed exports update existing records.
• X Article preservation — if the exporter includes article title/body/URL, that content is stored locally as an x-article link.
• Bounded enrichment — public HTTP(S) links only, redirect validation, response-size limits, timeouts, and content extraction.
• Docker + CI — non-root container, linting, tests, coverage, dependency audit, secret scan, private-data scan, and Docker validation.

Security Model

URL enrichment processes untrusted links, so the fetcher:

• accepts only http:// and https://;
• rejects loopback, private, link-local, multicast, reserved, and other non-public IP destinations;
• validates redirect targets before following them;
• limits redirects to five hops;
• streams responses with a configurable hard byte limit;
• accepts text/HTML-style responses only;
• applies request timeouts and a descriptive User-Agent.

This is defense in depth for a local enrichment tool. Do not expose the MCP server or its data directory directly to the public Internet.

Current Limits

• The project does not log into X or scrape authenticated X pages.
• Full X Article text is available only when the bookmark exporter includes that article content. A raw X Article URL by itself may not be enrichable without an authenticated browser/session.
• A bookmark marked as a reply can be identified as thread-related, but reconstructing an entire X conversation requires the exporter to include the conversation content.
• SQLite is intended for a personal/local collection, not multi-user hosted service use.

Quick Start

1. Export bookmarks

Export X bookmarks to JSON and save the file under data/exports/, for example:

data/exports/bookmarks.json

2. Install

git clone https://github.com/p0ki/x-bookmarks-mcp.git
cd x-bookmarks-mcp
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

3. Import and optionally enrich

python -m src import data/exports/bookmarks.json
python -m src enrich
python -m src stats

Re-running import is safe:

• new bookmark → added;
• changed bookmark/export metadata → updated;
• unchanged bookmark → skipped.

4. Run the MCP server

python -m src.server

Example MCP client configuration:

{
  "mcpServers": {
    "xbookmarks": {
      "command": "python",
      "args": ["-m", "src.server"],
      "cwd": "/absolute/path/to/x-bookmarks-mcp"
    }
  }
}

Docker

Build:

docker build -t xbookmarks-mcp .

Run the MCP server over stdio:

docker run --rm -i \
  -v "$PWD/data:/app/data" \
  -v "$PWD/config.yaml:/app/config.yaml:ro" \
  xbookmarks-mcp

CLI through Docker Compose:

docker compose run --rm xbookmarks-cli import data/exports/bookmarks.json
docker compose run --rm xbookmarks-cli enrich
docker compose run --rm xbookmarks-cli stats

How It Works

bookmark export
      |
      v
  ingest.py
      |
      v
 SQLite + FTS5 <---- notes / tags
      |
      +---- optional public URL enrichment
      |
      v
 MCP server (stdio)
      |
      v
 Claude / MCP client

1. Ingest detects supported JSON formats and updates the local collection.
2. Embedded article content is preserved when the export contains it.
3. Enrich fetches bounded public linked-page content and extracts readable text.
4. Tagger applies configurable keyword rules.
5. FTS5 indexes searchable fields; MCP writes rebuild the index immediately.
6. MCP exposes a small read/write tool contract.

MCP Tools

Configuration

Tag rules live in config.yaml.

Development

pytest -v
pytest --cov=src --cov-fail-under=80 -v

ruff check src/ tests/
black --check src/ tests/
isort --check-only src/ tests/

python scripts/check_private_data.py

CI tests Python 3.11, 3.12, and 3.13 and also performs dependency, secret, privacy, Docker-image, and Docker-Compose checks.

Dependency Policy

Runtime and development dependencies are pinned for reproducible CI/install behavior. The project intentionally remains on the supported MCP Python SDK 1.x line until a deliberate MCP v2 migration is completed.

License

MIT