rustpad-mcp

A Model Context Protocol (MCP) server for Rustpad, the efficient, minimal, self-hosted collaborative text editor.

Lets MCP clients like Claude Code, Claude Desktop or Codex read and write the pads of a Rustpad instance: fetch a document, create one, replace it wholesale or edit it in place.

Eight tools is the ceiling, not the floor: RUSTPAD_ALLOW_TOOLS=essential registers a curated five instead, and a model picks the right tool far more reliably from five than from eight — see choosing which tools load.

Reads go through Rustpad's HTTP API; writes speak the operational-transformation WebSocket protocol, so targeted edits (append_to_document, replace_in_document) merge cleanly with what human collaborators type at the same time instead of overwriting it. While the server edits a pad, it is visible to everyone in the pad as a collaborator named rustpad-mcp.

The two edits that cannot be undone ask a person. Where the client supports MCP elicitation, replacing a non-empty pad and search-replacing across more than one match raise a real dialog that the model cannot answer on its behalf — and the replace_in_document one says how many places are about to change. Where it does not, they fall back to a two-call token, and say so rather than implying somebody approved. ELICITATION=false takes that fallback deliberately; it never removes the guard. See Asking a person.

What makes it different

Real OT edits, not overwrites. append_to_document and replace_in_document retain everything they do not touch, and the Rustpad server transforms concurrent edits — a human typing in the same pad at the same moment loses nothing. The model shows up in the pad as a named collaborator.

Built for an unauthenticated world. Rustpad has no accounts, so every pad is untrusted by definition. Everything that comes out of one — reads, metadata, even upstream error bodies — is explicitly marked as data, never instructions, before a model sees it.

Related MCP server: Unofficial Overleaf MCP Server

Requirements

• A reachable Rustpad instance (self-hosted; the server is stateless and needs no credentials — Rustpad has no authentication)

• Node.js >= 24, or Docker

Configuration

The same URL serves the HTTP API, the WebSocket endpoint and the share links returned by the tools (<RUSTPAD_URL>/#<pad-id>). The RUSTPAD_* booleans must be exactly true. The server starts and lists its tools without configuration; every call then fails with setup instructions.

ELICITATION is the odd one out twice over: it carries no prefix, so it reaches every MCP server in the same environment, and a value that is neither true nor false stops the server rather than falling back — it is the only variable here that defaults to on, and a typo would otherwise leave the dialog running while you believed it was off. A server started with it off prints one line saying so.

Keep in mind what Rustpad is: pads are ephemeral (lost on server restart and after 24 hours of inactivity, unless the instance is run with SQLITE_URI) and anyone who knows a pad id can read and write it. Do not put secrets in pads.

Choosing which tools load

RUSTPAD_ALLOW_TOOLS and RUSTPAD_DENY_TOOLS take comma-separated tool names; a trailing * matches a whole family. essential is a curated preset of five: get_document, get_document_info, create_document, set_document, append_to_document.

RUSTPAD_ALLOW_TOOLS=essential
RUSTPAD_ALLOW_TOOLS=get_document,append_to_document
RUSTPAD_DENY_TOOLS=set_document

One boundary is softer than the list suggests: Rustpad has no create operation, a pad exists under any id the moment it is written to, so denying create_document does not stop new pads — set_document or append_to_document on a fresh id makes one just the same. Deny the write tools you mean, or use RUSTPAD_READ_ONLY.

An entry that matches no tool aborts startup and names it, so a typo cannot silently hide a tool — an absent tool is not something anyone traces back to an environment variable. A filtered tool is never registered, so it is absent from tools/list and unknown to tools/call alike, exactly like a write tool under RUSTPAD_READ_ONLY.

If you run several of these servers at once, mcp-hub is the other answer — its /hub endpoint replaces every server's tools with six meta-tools.

Installation

Claude Code

claude mcp add rustpad --env RUSTPAD_URL=https://rustpad.example.net -- npx rustpad-mcp

Claude Desktop

{
  "mcpServers": {
    "rustpad": {
      "command": "npx",
      "args": ["rustpad-mcp"],
      "env": {
        "RUSTPAD_URL": "https://rustpad.example.net"
      }
    }
  }
}

Codex

~/.codex/config.toml:

[mcp_servers.rustpad]
command = "npx"
args = ["-y", "rustpad-mcp"]

[mcp_servers.rustpad.env]
RUSTPAD_URL = "https://rustpad.example.net"

Docker

docker run -i --rm -e RUSTPAD_URL=https://rustpad.example.net ghcr.io/ni-c/rustpad-mcp

Through mcp-hub

A client that cannot spawn a local process — ChatGPT connectors, Claude on the web, Cursor, LibreChat — reaches rustpad-mcp through mcp-hub: one container serves many stdio MCP servers over Streamable HTTP, with an OAuth 2.1 login behind a single password and long-lived tokens for the clients that cannot do OAuth. Its /hub endpoint puts every server behind six meta-tools, so one connector reaches all of them without N×tool schemas in the model's context, and it speaks both protocol revisions — a question this server asks travels through it to the person at the far end.

Its /config/mcp.json uses Claude Code's format, so the entry is the one you already have:

{
  "mcpServers": {
    "rustpad": {
      "command": "npx",
      "args": ["-y", "rustpad-mcp"],
      "env": { "RUSTPAD_ALLOW_TOOLS": "essential" },
      "denyTools": ["set_document"]
    }
  }
}

allowTools and denyTools there are the hub's own per-server filter, which is not the same thing as *_ALLOW_TOOLS in env — the difference, and the mistake it invites, are in the client guide.

Tools

👤 asks a person through MCP elicitation · falls back to a two-call confirm_token where the client cannot show a dialog.

With RUSTPAD_READ_ONLY=true only the first three are registered.

Structured output

Every tool declares an outputSchema and answers with structuredContent alongside the text block, so a client can use the result without parsing prose. The five write tools used to answer with a sentence — "Appended 12 characters to pad …" — and the sentence is still there, in the text block:

{
  "id": "notes",
  "url": "https://rustpad.example/#notes",
  "appended_characters": 12,
  "characters": 137,
  "note": "Pads are ephemeral: …",
}

get_document answers {text} rather than the pad as the whole result, for the same reason get_document_info has always been an object: a schema whose root is a string is served to a 2025-era client rewritten as {result: …}, so the tool would answer in two shapes depending on who asked. It is also where empty and truncated can live — an empty answer used to be a sentence.

The two read tools that report pad content carry untrusted: true and source: "rustpad" as fields. A pad is world-writable to anyone who knows its id, including text this server wrote earlier, and a client that reads the structured half would otherwise get it with no framing at all.

Not exposed, on purpose

No pad listing — Rustpad has no such API. Pads exist implicitly under every id, so you have to know the ids you care about. get_stats reports how many documents the server currently holds, but not their names.

No accounts, no permissions. Rustpad has neither, which is why every pad is treated as untrusted input rather than as something a login vouched for.

Safety

• Pad content is world-writable and therefore untrusted: every read result is prefixed with a marker telling the model to treat it as data, never as instructions.

• The two irreversible edits ask a person: a real dialog the model cannot answer on its behalf, bound to the pad and the exact replacement. Where the client cannot show one, a single-use token that only ever appears in a previous tool result — which proves the call was made twice with the same arguments, and nothing more. The fallback text says which of the two it was.

• Tool results are size-capped; control characters are stripped from everything the instance wrote, and upstream error bodies are sanitized before they reach the model.

• The instance is untrusted too: WebSocket frames are limited at the header, before they are buffered, the message queue is bounded in bytes, and every frame is checked for shape before it is folded into the pad.

• RUSTPAD_INSECURE_TLS relaxes certificate validation only for the configured connection, never process-wide.

Documentation

The full guide, tool reference and security notes live at rustpad-mcp.ni-c.de (source in docs/).

Development

npm install
npm run lint && npm run build && npm test

The test suite talks to an in-memory fake of rustpad-server (including OT transformation of concurrent edits) over the real MCP protocol; no live instance is needed. The architecture diagram and social card are generated — edit docs/assets/architecture.source.svg and run npm run assets, never the rendered copies.

Releasing

Releases are tag-driven. Bump package.json, move the [Unreleased] notes in CHANGELOG.md under the new version, commit, then:

git tag -s vX.Y.Z -m "vX.Y.Z"
git push origin main vX.Y.Z

The release workflow publishes to npm via Trusted Publishing (OIDC, with provenance), pushes the multi-arch container image to GHCR, creates the GitHub release from the CHANGELOG section, and updates the entry in the official MCP registry.

Contributing

Issues, discussions and pull requests are welcome — see CONTRIBUTING.md. For vulnerabilities please use private reporting rather than a public issue; the policy is in SECURITY.md.

License

MIT © Willi Thiel