lesswrong-mcp

A read-only MCP server for LessWrong and the Alignment Forum. It lets a model search the two forums and read what's there: a post and its comment thread, a user's profile, a tag's wiki page, a sequence's ordered posts, a front-page feed, or a filtered slice of posts by author, karma, and date. There's no API key and no login involved, since it only reads the sites' public endpoints.

LessWrong and the Alignment Forum run on the same ForumMagnum codebase and share a database, so this one server covers both, and search spans both at once. The Alignment Forum is really a curated slice of LessWrong rather than a separate site, so everything on it is on LessWrong too. The other tools take an optional site (lesswrong by default, or alignmentforum), which mostly changes the feeds and the karma you see.

The tools

Tool	What it does
lw_search	Full-text search over posts, comments, tags, users, and sequences, across both forums (pass type to search just one kind)
lw_get_post	A post's body and metadata, by id or slug
lw_get_comments	A post's comment thread, sorted by top, new, or old
lw_get_user	Someone's profile and their best posts
lw_get_tag	A tag or wiki page and the posts filed under it
lw_get_sequence	A sequence's intro and its ordered list of posts
lw_list_feed	A front-page feed: latest, recent, curated, or home
lw_filter_posts	Posts filtered by author, tag, date range, and karma, with a sort order

Most of these just wrap the forum's Markdown API, which the site publishes for exactly this kind of use, so the responses come back as clean Markdown already. lw_filter_posts is the one exception: it goes through GraphQL, because that's the only way to get the precise author + tag + date + karma filtering.

Installing it

Two ways to run it, depending on where you use Claude.

In the desktop app

Grab the bundle: lesswrong-mcp.mcpb. That link always points at the newest build, or you can pick it up from the latest release.

In Claude Desktop, go to Settings → Extensions → Advanced settings → Install Extension, and choose the file. All eight tools show up right away. Everything runs locally, and the bundle uses the uv runtime, so Claude Desktop takes care of Python and the dependencies for you.

As a remote connector

For claude.ai on the web, on mobile, in Cowork, or through the API, you host the server yourself and add it as a custom connector. There's a Dockerfile in the repo, so you can deploy it wherever you like: Render, Fly.io, Railway, a plain VPS, or an MCP host such as Manufact. Run it in HTTP mode with MCP_TRANSPORT=http and MCP_HOST=0.0.0.0, and it'll use whatever $PORT the platform gives it. The MCP endpoint is at /mcp.

Then, in claude.ai, go to Customize → Connectors → Add custom connector, paste https://<your-host>/mcp, and add it. No OAuth needed.

Running it locally

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

python lesswrong_mcp.py                                  # stdio (Claude Desktop / MCP Inspector)
MCP_TRANSPORT=http MCP_PORT=8000 python lesswrong_mcp.py # HTTP, endpoint at /mcp

The environment variables it reads: MCP_TRANSPORT (stdio or http), MCP_HOST (127.0.0.1 by default, 0.0.0.0 when you're hosting it), and MCP_PORT / PORT (8000 by default). Two optional ones harden the HTTP transport: set MCP_ALLOWED_HOSTS and/or MCP_ALLOWED_ORIGINS (comma-separated) to turn on DNS-rebinding (Host/Origin) protection scoped to those values — useful for a deployment with a known hostname. Without them, protection is still on automatically for a loopback bind (scoped to localhost:<port>, covering the local HTTP example above); it's only disabled for a non-loopback bind like 0.0.0.0, whose public hostname a PaaS assigns at deploy time. In HTTP mode the server runs statelessly (no per-session affinity), so it works behind a load balancer or a scale-to-zero host.

Tests

pip install -e '.[dev]'   # or: pip install pytest
pytest                    # hermetic: mocks the HTTP/GraphQL layer, no network
LW_LIVE_TESTS=1 pytest    # also runs the one live-endpoint test against the real API

The suite lives in tests/ and is hermetic — it stubs the HTTP/GraphQL layer and drives the async tool handlers directly, so it runs offline and fast. It covers the shared HTTP layer (retry/backoff, Retry-After, the overall deadline, and the SSRF redirect guard), the GraphQL plumbing (error surfacing and slug→id resolution), each tool's routing and paging, and the selector that lw_filter_posts builds — in particular that a tag rides on the requested sort view as a filterSettings "Required" tag, composing with any sort order and the other filters.

Building the bundle

mkdir -p dist
npx @anthropic-ai/mcpb pack . dist/lesswrong-mcp.mcpb

manifest.json and pyproject.toml describe the bundle, and .mcpbignore keeps the remote-only files (the Dockerfile, requirements.txt, and so on) out of it.

The bundle ships unsigned. Don't run mcpb sign --self-signed on it: mcpb appends its signature block after the zip's end-of-central-directory record without recording it as a zip comment, so strict zip parsers — including the desktop extension installer — reject the signed file (Invalid comment length … extra bytes at the end). Self-signing also only yields an "unverified publisher" bundle, so it buys nothing here. If you ever need a genuinely trusted signature, use a CA-issued certificate (mcpb sign --cert/--key) and confirm the result still installs.

A few things worth knowing

• It's read-only. The forum API can also create and edit drafts, but those endpoints are deliberately left out here.
• Pagination works a little differently per tool: lw_search takes page and limit, lw_filter_posts takes limit and offset, and lw_get_comments takes limit (a comment count, up to 2000, since some threads run to hundreds of comments). lw_get_post and lw_get_comments both return everything by default, but a long one can be capped with max_chars and paged with offset — when a slice stops short, the response ends with the exact offset to pass next. The home feed is the one feed that ignores limit: it's a fixed composite page, so reach for latest/recent/curated when you want a set count.
• Comments come back most complete from the default lesswrong site, which returns the whole thread. The alignmentforum view of a post only shows the comments promoted to the Alignment Forum, so it's a subset.
• Search always covers both forums and can't be narrowed to one, since they share a single index. If you specifically want Alignment-Forum-promoted posts, reach for lw_filter_posts with alignment_forum_only: true.
• It tries not to be rude to the forum: in-flight requests are capped, and it backs off and retries on 429s and 5xx errors.
• The remote version only ever touches public data and can't write anything, but if you put it on a public URL you can still limit inbound traffic to Anthropic's IP ranges or add OAuth in front of it.

License

MIT