Linkwarden

# linkwarden-mcp **MCP server for self-hosted [Linkwarden](https://linkwarden.app/): ask your AI about bookmarks, collections, and tags.** [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE) [![Python](https://img.shields.io/badge/python-%3E%3D3.10-blue.svg)](https://www.python.org/) [![MCP](https://img.shields.io/badge/MCP-stdio-green.svg)](https://modelcontextprotocol.io/) [![CI](https://github.com/flumpiey/linkwarden-mcp/actions/workflows/ci.yml/badge.svg)](https://github.com/flumpiey/linkwarden-mcp/actions/workflows/ci.yml) [![PyPI](https://img.shields.io/pypi/v/linkwarden-mcp.svg)](https://pypi.org/project/linkwarden-mcp/) ## What is Linkwarden? [Linkwarden](https://linkwarden.app/) is a self-hosted, open-source bookmark manager. You collect, organize, annotate, and preserve webpages in one place, with full-page archives so content stays readable after the original page disappears. It also supports collaboration and public sharing. This project wires the Linkwarden HTTP API into the [Model Context Protocol](https://modelcontextprotocol.io/) so Cursor, Claude, VS Code Copilot, and other MCP hosts can query your live library in natural language. Useful Linkwarden links: - [Documentation](https://docs.linkwarden.app/) - [GitHub](https://github.com/linkwarden/linkwarden) - [Cloud](https://linkwarden.app/) - [Demo](https://demo.linkwarden.app/) - [Self-hosting setup](https://docs.linkwarden.app/self-hosting/installation) ## What this server does Default is **read-only**. You get: - **18 read tools** - discovery (`list_resources`), six core reads (search, get, preserved content, collections, tags, overview), and eleven triage/hygiene workflows - **Task tools (opt-in)** - intent-shaped writes such as `save_link`, `smart_save_link`, `organise_links`, `create_collection`, `apply_triage_plan` (register when matching write scopes are set) - **Delete tools (opt-in)** - `delete_links`, `delete_tags`, `merge_tags`, `delete_collection` (register only under delete scopes; never implied by write) - **Hard denylist** - tokens, session, auth, user admin (except `GET /api/v1/users/me`), migration, and whole-instance preservation stay blocked even when writes are on Transport is **stdio**. No HTTP server. No global install required if you use [`uv`](https://docs.astral.sh/uv/) / `uvx`. ## Branding / icons Four surfaces (keep them in sync when the mark changes): 1. **stdio hosts (Cursor, Claude Desktop via `mcp.json`):** `serverInfo.icons` from [`server_icons()`](src/linkwarden_mcp/server.py) — embedded data URI from [`src/linkwarden_mcp/assets/icon.png`](src/linkwarden_mcp/assets/icon.png), plus HTTPS fallback [`docs/icon-512.png`](docs/icon-512.png) (`https://raw.githubusercontent.com/flumpiey/linkwarden-mcp/main/docs/icon-512.png`). `website_url` is `https://linkwarden.app/`. 2. **Cursor plugin:** [`.cursor-plugin/plugin.json`](.cursor-plugin/plugin.json) `logo` → [`docs/linkwarden-icon.svg`](docs/linkwarden-icon.svg). 3. **Claude Desktop Extension:** [`mcpb/icon.png`](mcpb/icon.png) (packed with `npx @anthropic-ai/mcpb pack mcpb`). 4. **Claude.ai remote connectors:** Claude.ai ignores `serverInfo.icons` and uses the **root-domain favicon** of the connector URL. If you host a remote MCP later, serve [`docs/favicon.ico`](docs/favicon.ico) at the registrable domain root (e.g. `https://acme.com/favicon.ico` for `https://mcp.acme.com/...`). [`server.json`](server.json) registry metadata also points its `icons[0].src` at the same raw `docs/icon-512.png` URL. ## Requirements - Python ≥ 3.10 (pulled in automatically by `uvx`) - [uv](https://docs.astral.sh/uv/) (provides `uvx`) - A reachable Linkwarden instance: `LINKWARDEN_API_URL` + `LINKWARDEN_API_KEY` ### Access token 1. Sign in to your Linkwarden instance (self-hosted or Cloud). 2. Open **Settings → Access Tokens** (or go to `/settings/access-tokens`). 3. Create a **New Access Token**, give it a name, and copy the value into `LINKWARDEN_API_KEY`. 4. Set `LINKWARDEN_API_URL` to your instance base URL (usually without `/api/v1`; include `/api/v1` only if your deployment requires it), e.g. `https://links.example.com` or local Docker `http://127.0.0.1:3000`. `linkwarden-mcp` sends the token as `Authorization: Bearer …`. API overview: [API Introduction](https://docs.linkwarden.app/api/api-introduction). Copy [`.env.example`](.env.example) to `.env` for local runs — **never commit `.env`**. Prefer the Cursor plugin **Configure** UI for credentials, or a secret manager in production. ## Quick start Run the [PyPI](https://pypi.org/project/linkwarden-mcp/) package with [`uvx`](https://docs.astral.sh/uv/guides/tools/): ```bash uvx linkwarden-mcp ``` Paste a client config below, set `LINKWARDEN_API_URL` / `LINKWARDEN_API_KEY`, restart the host, then ask: *“Find my unread bookmarks about Python”* or *“What's in my Dev collection?”* From a git clone (dev): `uvx --from git+https://github.com/flumpiey/linkwarden-mcp linkwarden-mcp` or `uv run --directory /path/to/linkwarden-mcp linkwarden-mcp`. ## Installation Configs below pull [`linkwarden-mcp`](https://pypi.org/project/linkwarden-mcp/) from PyPI. Leave write-scope env vars unset for read-only.
Cursor **Plugin (Configure UI for URL, key, and scopes):** this repo is a Cursor plugin via [`.cursor-plugin/plugin.json`](.cursor-plugin/plugin.json) + root [`mcp.json`](mcp.json). 1. Symlink or copy the clone to `~/.cursor/plugins/local/linkwarden-mcp` (Windows: `%USERPROFILE%\.cursor\plugins\local\linkwarden-mcp`). - **macOS / Linux:** `ln -s /path/to/linkwarden-mcp ~/.cursor/plugins/local/linkwarden-mcp` - **Windows:** Cursor does **not** follow symlinks for local plugins. Use a junction or copy instead: ```bat mklink /J "%USERPROFILE%\.cursor\plugins\local\linkwarden-mcp" "E:\Development\linkwarden-mcp" ``` or: ```bat robocopy "E:\Development\linkwarden-mcp" "%USERPROFILE%\.cursor\plugins\local\linkwarden-mcp" /E ``` 2. Reload the window. 3. Open **Plugins → Configure** on `linkwarden-mcp`. Set **Linkwarden API URL** and **Linkwarden API key**. Leave **Write scopes** / **Delete scopes** empty for read-only, or paste a CSV such as `links,collections`. 4. Confirm the `linkwarden` MCP server is enabled under Customize / MCP. Marketplace listing is a separate submit at [cursor.com/marketplace/publish](https://cursor.com/marketplace/publish). **Manual `mcp.json`:** project [`.cursor/mcp.json`](.cursor/mcp.json) or user-wide `~/.cursor/mcp.json`. Root [`mcp.json`](mcp.json) is plugin wiring with `${…}` placeholders only — never commit real secrets there. From PyPI: ```json { "mcpServers": { "linkwarden": { "type": "stdio", "command": "uvx", "args": ["linkwarden-mcp"], "env": { "LINKWARDEN_API_URL": "https://links.example.com", "LINKWARDEN_API_KEY": "your-token" } } } } ``` Local editable (dev): ```json { "mcpServers": { "linkwarden": { "type": "stdio", "command": "uv", "args": ["run", "--directory", "/path/to/linkwarden-mcp", "linkwarden-mcp"], "env": { "LINKWARDEN_API_URL": "https://links.example.com", "LINKWARDEN_API_KEY": "your-token" } } } } ``` Optional scoped writes in the `env` block: ```json "LINKWARDEN_MCP_WRITE_SCOPES": "links,collections", "LINKWARDEN_MCP_DELETE_SCOPES": "links" ``` Restart Cursor after saving. Confirm `linkwarden` under MCP settings.
Claude Desktop **Desktop Extension (`.mcpb`):** download [`mcpb.mcpb`](https://github.com/flumpiey/linkwarden-mcp/releases/latest/download/mcpb.mcpb) from [GitHub Releases](https://github.com/flumpiey/linkwarden-mcp/releases). Use **v0.1.5+** (needs [`uv`](https://docs.astral.sh/uv/) on PATH). Launch is `uv tool run --python 3.12 linkwarden-mcp`. Do not put the PyPI package in `mcpb/pyproject.toml` dependencies — Claude Desktop syncs that file at install and can fail on system Python 3.13. 1. Open Claude Desktop → **Settings → Extensions**. 2. Open **Advanced settings** → **Install Extension…** 3. Select `mcpb.mcpb`. Review permissions, enter **Linkwarden API URL** and **Linkwarden API key**, then click **Install**. 4. Leave **Write scopes** and **Delete scopes** empty for read-only. 5. Restart Claude Desktop if tools do not appear. Build your own bundle from a clone: ```bash npx @anthropic-ai/mcpb pack mcpb ``` On Windows, double-click often does nothing and dragging the file into chat attaches it to the conversation instead of installing it. Use **Install Extension…** in Settings. **Manual `claude_desktop_config.json` fallback:** edit the Claude Desktop config, then restart the app. | OS | Path | |----|------| | macOS | `~/Library/Application Support/Claude/claude_desktop_config.json` | | Windows | `%APPDATA%\Claude\claude_desktop_config.json` | ```json { "mcpServers": { "linkwarden": { "command": "uvx", "args": ["linkwarden-mcp"], "env": { "LINKWARDEN_API_URL": "https://links.example.com", "LINKWARDEN_API_KEY": "your-token" } } } } ``` Local clone: ```json { "mcpServers": { "linkwarden": { "command": "uv", "args": ["run", "--directory", "/path/to/linkwarden-mcp", "linkwarden-mcp"], "env": { "LINKWARDEN_API_URL": "https://links.example.com", "LINKWARDEN_API_KEY": "your-token" } } } } ```
Claude Code Add via CLI: ```bash claude mcp add linkwarden --env LINKWARDEN_API_URL=https://links.example.com --env LINKWARDEN_API_KEY=your-token -- uvx linkwarden-mcp ``` Or edit `~/.claude.json` / project MCP config: ```json { "mcpServers": { "linkwarden": { "command": "uvx", "args": ["linkwarden-mcp"], "env": { "LINKWARDEN_API_URL": "https://links.example.com", "LINKWARDEN_API_KEY": "your-token" } } } } ```
VS Code / GitHub Copilot Create [`.vscode/mcp.json`](.vscode/mcp.json) in the project root: ```json { "servers": { "linkwarden": { "type": "stdio", "command": "uvx", "args": ["linkwarden-mcp"], "env": { "LINKWARDEN_API_URL": "https://links.example.com", "LINKWARDEN_API_KEY": "your-token" } } } } ``` Local editable: ```json { "servers": { "linkwarden": { "type": "stdio", "command": "uv", "args": ["run", "--directory", "/path/to/linkwarden-mcp", "linkwarden-mcp"], "env": { "LINKWARDEN_API_URL": "https://links.example.com", "LINKWARDEN_API_KEY": "your-token" } } } } ``` Reload the window. Open Copilot Chat and confirm the `linkwarden` tools are available.
Windsurf Edit `~/.codeium/windsurf/mcp_config.json` (macOS/Linux) or the Windsurf MCP settings UI: ```json { "mcpServers": { "linkwarden": { "command": "uvx", "args": ["linkwarden-mcp"], "env": { "LINKWARDEN_API_URL": "https://links.example.com", "LINKWARDEN_API_KEY": "your-token" } } } } ``` Restart Windsurf after saving.
Zed Add under `context_servers` in Zed `settings.json` (Agent Panel → settings also works): ```json { "context_servers": { "linkwarden": { "command": "uvx", "args": ["linkwarden-mcp"], "env": { "LINKWARDEN_API_URL": "https://links.example.com", "LINKWARDEN_API_KEY": "your-token" } } } } ```
Cline Edit the Cline MCP settings file (`cline_mcp_settings.json` via the Cline MCP UI): ```json { "mcpServers": { "linkwarden": { "command": "uvx", "args": ["linkwarden-mcp"], "env": { "LINKWARDEN_API_URL": "https://links.example.com", "LINKWARDEN_API_KEY": "your-token" } } } } ```
Continue In `.continue/config.yaml`: ```yaml mcpServers: - name: linkwarden command: uvx args: - linkwarden-mcp env: LINKWARDEN_API_URL: https://links.example.com LINKWARDEN_API_KEY: your-token ```
Generic / any stdio MCP host Any host that can spawn a stdio MCP server: | Field | Value | |-------|-------| | Command | `uvx` | | Args | `linkwarden-mcp` | | Env | `LINKWARDEN_API_URL`, `LINKWARDEN_API_KEY` (+ optional write scopes) | ```bash uvx linkwarden-mcp ``` Dev from a clone: `uv run --directory /path/to/linkwarden-mcp linkwarden-mcp`. `npx` only runs npm packages. This is a Python package; use `uvx`.
## Environment | Variable | Required | Notes | |----------|----------|-------| | `LINKWARDEN_API_URL` | yes | Base URL (include `/api/v1` only if required; typical: `https://links.example.com`) | | `LINKWARDEN_API_KEY` | yes | Access token from Settings → Access Tokens; sent as `Authorization: Bearer`; never logged | | `LINKWARDEN_MCP_WRITE_SCOPES` | no | Comma-separated domains for create/update. Empty = no writes. | | `LINKWARDEN_MCP_DELETE_SCOPES` | no | Comma-separated domains for delete only. Never implied by WRITE_SCOPES. | | `LINKWARDEN_MAX_BULK` | no | Max records per bulk op (default `25`) | | `TEST_LINKWARDEN_API_URL` | integration only | Live sandbox URL for `pytest -m integration` | | `TEST_LINKWARDEN_API_KEY` | integration only | Live sandbox token for `pytest -m integration` | Valid scopes: `links`, `collections`, `tags`, `raw`. No wildcards (`*`, `all`). `raw` expands effective scopes to all domain scopes (escape hatch). **Recommended** (covers most bookmark workflows without every mutating tool): ```json "LINKWARDEN_MCP_WRITE_SCOPES": "links,collections", "LINKWARDEN_MCP_DELETE_SCOPES": "links" ``` Default with no scopes: **18 tools**. All three domain scopes in WRITE and DELETE: **31 tools**. Legacy `LINKWARDEN_MCP_ALLOW_WRITES` / `ALLOW_WRITES` / `LINKWARDEN_MCP_WRITES` hard-fail if set. Use the scoped vars instead. MCP host env (`.cursor/mcp.json` or Cursor plugin Configure) **must match** process env / `.env` or scope behavior drifts. See [`.env.example`](.env.example). **Never commit `.env`.** Prefer Cursor plugin Configure UI or a secret manager in production. ## Write scopes and task tools When a scope is listed in `LINKWARDEN_MCP_WRITE_SCOPES`, the server registers **task tools** for that domain. `LINKWARDEN_MCP_DELETE_SCOPES` enables delete/merge tools per domain. Call `list_resources` to inspect `read_only`, scope lists, and the live boundary string. | Tool | Scopes | Purpose | |------|--------|---------| | `save_link` | WRITE `links` | Save a URL into a collection (by name) | | `smart_save_link` | WRITE `links` | Save with optional heuristic collection/tags | | `organise_links` | WRITE `links` | Move or retag multiple links | | `update_link` | WRITE `links` | Update link fields (read-modify-write) | | `queue_archive` | WRITE `links` | Queue preservation (async; not immediate) | | `apply_triage_plan` | WRITE `links` | Apply `[{link_id, collection?, tags?}]`; default `dry_run=true` | | `bulk_sort_by_rules` | WRITE `links` | Match `domain_pattern` rules then organise; default `dry_run=true` | | `create_collection` | WRITE `collections` | Create a collection (optional parent) | | `auto_tag_by_domain` | WRITE `links` + `tags` | Apply domain→tag rules; default `dry_run=true` | | `delete_links` | DELETE `links` | Delete multiple links | | `delete_tags` | DELETE `tags` | Delete tags by id or name | | `merge_tags` | DELETE `tags` | Merge tags into a new name (destructive) | | `delete_collection` | DELETE `collections` | Delete a collection after user chooses delete/move/cancel for its links (elicitation or `on_links`) | Example with recommended scopes only: ```json "LINKWARDEN_MCP_WRITE_SCOPES": "links,collections", "LINKWARDEN_MCP_DELETE_SCOPES": "links" ``` **Denylist (always blocked):** `/api/v1/tokens`, `/api/v1/session`, `/api/v1/auth`, `/api/v1/users/**` (except `GET /api/v1/users/me`), migration, and whole-instance preservation worker actions. ## Tools ### Read tools Always registered (18 total). | Tool | Purpose | |------|---------| | `list_resources` | Discovery; reports `read_only` + live write/delete scopes | | `search_links` | Search by query, collection, tag, or pin status | | `get_link` | Full metadata for one link | | `read_link_content` | Preserved plain text (`textContent` or archive fallback) | | `list_collections` | Collections with link counts | | `list_tags` | Tags with link counts | | `get_library_overview` | Totals, empty collections, unused tags | | `suggest_collection_for_url` | Heuristic collection suggestions for a URL | | `suggest_tags_for_link` | Suggest existing-library tags (never invents names) | | `find_unsorted_links` | List unsorted links (default collection: Unorganized) | | `triage_links` | Propose collection/tags for link ids (no writes) | | `find_duplicate_links` | Group links with the same normalized URL | | `recommend_collection_for_links` | Consensus collection for a batch of links | | `suggest_links_for_collection` | Find links elsewhere that likely belong | | `analyze_collection_overlap` | Compare two collections for shared domains/tags/URLs | | `suggest_collection_structure` | Hygiene: empty, near-duplicate names, overcrowded | | `align_tags_with_similar_links` | Tags used on similar-domain links | | `get_sorting_dashboard` | One-shot triage: unsorted, duplicates, empty, largest | ### Write tools Registered only when matching scopes are set (see table above). Prefer `smart_save_link` / triage tools over raw field edits when you are sorting an inbox. | Pattern | Requires | Notes | |---------|----------|-------| | Link create/update/organise/archive | WRITE `links` | Includes workflow writers with `dry_run` defaults | | Collection create | WRITE `collections` | Optional parent by name | | Domain auto-tag | WRITE `links` + `tags` | Only existing tag names | | Deletes / tag merge | matching DELETE scope | Destructive; confirm ids first | ## Agent Skill Companion skill: [`skills/linkwarden-bookmarks/SKILL.md`](skills/linkwarden-bookmarks/SKILL.md). The Cursor plugin discovers this skill from `skills/`. Without the plugin, copy or symlink that folder into your agent skills path. It tells the model to call `list_resources` first, verify after writes, and which workflow tools to prefer. ## Development ```bash uv sync --extra dev npm install # installs lefthook + commitlint; registers git hooks uv run linkwarden-mcp ``` Offline tests only (respx). No live Linkwarden required: ```bash uv run ruff check src tests uv run pytest ``` ### Git hooks (lefthook) | Hook / command | What it runs | |----------------|--------------| | **pre-commit** | `ruff check` + full `pytest` (mocked suite) | | **commit-msg** | [commitlint](https://commitlint.js.org/) Conventional Commits | | **`npm run pre-publish`** | ensure build/twine → ruff → pytest → sdist contents → `python -m build` → `twine check` → `npx @anthropic-ai/mcpb pack mcpb` | Commit messages must follow Conventional Commits, e.g. `feat(api): add delete_links tool`. Run the local publish gate before tagging a release: ```bash npm run pre-publish ``` GitHub Actions matrix: Python 3.10 and 3.12. ## Caveats - One process ↔ one `LINKWARDEN_API_URL`. Multi-instance routing is out of scope. - Multi-user / team disambiguation on a shared instance is **unverified**. Do not claim multi-tenant support until validated against a live shared library. - Collection/tag suggestions are heuristic and library-local; they do not invent new tag names. - Bulk mutating workflows default to `dry_run=true`; set `dry_run=false` only after you review the plan. - ChatGPT Apps need a hosted HTTP MCP endpoint. This package is stdio-only. ## License MIT. See [LICENSE](LICENSE).