AFFiNE MCP Server

A Model Context Protocol (MCP) server for AFFiNE. It exposes AFFiNE workspaces and documents to AI assistants over stdio (default) or HTTP (/mcp) and supports both AFFiNE Cloud and self-hosted deployments.

Table of Contents

• Overview

• Choose Your Path

• Quick Start

• Compatibility Matrix

• Tool Surface

• Documentation Map

• Verify Your Setup

• Security and Scope

• Development

• Release Notes

• License

• Support

Related MCP server: mcp-server-ispace-wiki

Overview

AFFiNE MCP Server is designed for three common scenarios:

• Run a local stdio MCP server for Claude Code, Codex CLI, Cursor, or Claude Desktop

• Expose a remote HTTP MCP endpoint for hosted or browser-connected clients

• Automate AFFiNE workspace, document, database, organization, and comment workflows through a stable MCP tool surface

Highlights:

• Supports AFFiNE Cloud and self-hosted AFFiNE instances

• Supports stdio and HTTP transports

• Coordinates concurrent writes per workspace through one shared MCP server; optional document revisions reject stale edits

• Supports session-cookie and email/password authentication, plus compatible bearer tokens for older deployments

• Exposes 112 canonical MCP tools backed by AFFiNE GraphQL, WebSocket, and REST APIs

• Reads native database created-by (user ID) and created-time (timestamp) columns from row metadata; missing or invalid metadata returns null without backfilling it

• Creates and updates code blocks with optional native preview state, with language and preview readback

• Places, resizes, and connects images on edgeless canvases, with blob identifiers and geometry in readback

• Appends canvas frames in creation-order presentation sequence while preserving their native background layering

• Sets and clears native AFFiNE Journal dates with set_doc_journal

• Includes semantic page composition, native template instantiation, database intent composition, capability and fidelity reporting, and workspace blueprint helpers

• Includes Docker images, health probes, and end-to-end test coverage

Scope boundaries:

• This server can access only server-backed AFFiNE workspaces

• Browser-local workspaces stored only in local storage are not available through AFFiNE server APIs

• After a content write, transient workspace timestamp failures are retried up to three times with backoff. If metadata still cannot be confirmed, inspect the saved document before any repair; do not repeat the content edit. See document mutation recovery.

• AFFiNE 0.27+ removed the legacy personal-access-token GraphQL API; this server no longer exposes token-management tools

• AFFiNE Cloud requires browser-session authentication for this external GraphQL integration; programmatic email/password sign-in is blocked by Cloudflare

• delete_doc removes workspace page metadata and requests content deletion. Some AFFiNE servers retain content despite acknowledging the request; the tool returns a partial result with a warning unless content absence is verified. Use trash_doc for recoverable cleanup.

New in v3.2.1: Scripted cookie login now keeps session secrets out of process arguments, validates workspace access before saving credentials, and restores document pagination for ordinary workspace members.

Choose Your Path

Quick Start

1. Install the CLI

npm i -g affine-mcp-server
affine-mcp --version

You can also run the package ad hoc:

npx -y -p affine-mcp-server affine-mcp -- --version

2. Or run the server in Docker

docker run -d \
  -p 3000:3000 \
  -e MCP_TRANSPORT=http \
  -e AFFINE_BASE_URL=https://your-affine-instance.com \
  -e AFFINE_EMAIL=you@example.com \
  -e AFFINE_PASSWORD=your-password \
  -e AFFINE_MCP_AUTH_MODE=bearer \
  -e AFFINE_MCP_HTTP_TOKEN=your-strong-secret \
  ghcr.io/dawncr0w/affine-mcp-server:3.13.0

Then point your client at:

{
  "mcpServers": {
    "affine": {
      "type": "http",
      "url": "http://localhost:3000/mcp",
      "headers": {
        "Authorization": "Bearer your-strong-secret"
      }
    }
  }
}

For Docker, health checks, and remote deployment details, see docs/configuration-and-deployment.md#docker.

HTTP deployments have two environment-only session limits:

Short-lived Streamable HTTP clients, including cron jobs, should send DELETE /mcp with their Mcp-Session-Id and authentication headers when finished. Exiting the client process alone does not terminate its server-side session. A long idle timeout can let abandoned sessions fill the limit and cause 503 / -32002 errors. Close unused sessions and choose an appropriate idle timeout before raising the cap. See session capacity troubleshooting.

3. Save credentials with interactive login

affine-mcp login

This stores credentials in $XDG_CONFIG_HOME/affine-mcp/config when XDG_CONFIG_HOME is set, otherwise in ~/.config/affine-mcp/config, with mode 600.

The login flow reuses the configured or saved AFFiNE URL when you press Enter at the URL prompt. After authentication, workspace discovery displays the workspace name first and falls back to Workspace name unavailable when profile metadata has no name; the full ID remains visible. An invalid numeric selection is rejected and prompted again. Enter q or send end-of-file to cancel without writing a new config. A failed discovery request is reported as a failure; it is not treated as a completed login.

• For AFFiNE Cloud, paste the Cookie request header from a signed-in browser session

• For self-hosted AFFiNE, use email/password (recommended) or a signed-in session cookie

• AFFINE_API_TOKEN remains available only for deployments that still accept a compatible GraphQL bearer token

The prompt defaults to AFFINE_BASE_URL from the environment or the saved config file, so pressing Enter keeps an already-configured self-hosted URL instead of switching back to AFFiNE Cloud.

For a self-hosted instance reached over plain HTTP on a trusted private network, AFFINE_ALLOW_INSECURE_HTTP=true must be set for the login run as well as for the server. The opt-in is read from the environment first and then from the saved config file.

To avoid re-running login when a session expires, persist the account credentials instead of the session cookie:

affine-mcp login --save-credentials

With the email/password method, this stores AFFINE_EMAIL and AFFINE_PASSWORD so the server signs in on its own and renews the session before it expires. The password is written to the mode-600 config file, so use a dedicated least-privilege AFFiNE account. Without this flag the CLI keeps storing only the session credential, which never renews by itself.

For scripted session-cookie setup, keep the cookie out of process arguments:

affine-mcp login --url https://app.affine.pro --cookie-stdin --workspace-id your-workspace-id --force

Paste the cookie at the hidden prompt, or pipe it from a trusted secret source. The CLI verifies --workspace-id against the authenticated account before saving it. Piped input requires --force when existing credentials would be replaced.

After login, inspect and switch the saved default workspace without signing in again:

affine-mcp workspaces
affine-mcp workspaces --json
affine-mcp workspace <workspace-id>

workspaces lists names first and marks the current default. workspace validates the requested ID with the active credentials and changes only the local default (AFFINE_WORKSPACE_ID); it does not grant access to a workspace or change the authenticated account. With no ID, it uses the same validated selection flow.

4. Register the server with your client

Claude Code project config:

{
  "mcpServers": {
    "affine": {
      "command": "affine-mcp"
    }
  }
}

Codex CLI:

codex mcp add affine -- affine-mcp

More client-specific setup is in docs/client-setup.md.

5. Verify the connection

affine-mcp status
affine-mcp doctor
affine-mcp snippet codex

The recommended Codex form is a command-only registration line that keeps using the current saved login. snippet claude and snippet cursor print JSON configuration instead. Apply the generated snippet, then restart or reconnect the MCP client so it starts a fresh server process.

For an explicit environment snapshot, add --env to any snippet command. It copies the currently resolved URL, authentication, headers, workspace, and relevant OAuth settings; client environment variables win over saved config at runtime. Remove or regenerate copied credentials when they expire instead of expecting a saved login to override them.

If you want to expose the server remotely over HTTP instead of stdio, start with docs/configuration-and-deployment.md. If an HTTP server already runs on the same host as your stdio client, use the private stdio HTTP bridge instead of starting another full server process.

Compatibility Matrix

Node.js 20.18.1 is the minimum supported runtime. CI validates the Node.js 20, 22, 24, and 26 release lines.

Tool Surface

tool-manifest.json is the source of truth for canonical tool names. The MCP server exposes those tools through tools/list and tools/call; tool definitions returned by tools/list include MCP annotations that mark read-only, destructive, idempotent, and external-world behavior for client-side tool selection.

Every canonical tool also declares an MCP outputSchema for its structuredContent. Object results retain their existing top-level fields, while array and scalar results use stable { items }, { text }, or { value } envelopes. The existing text content remains unchanged for compatibility with clients that do not consume structured results.

get_capabilities exposes the full implemented list in server.supportedTools and the currently enabled surface in server.effective.profile and server.effective.enabledTools, after profiles, disabled groups or tools, and auth-mode policy. Inspect tools/list when you need the final set of callable tools.

Advertised input and output schemas omit the SDK-generated draft-07 $schema marker. Schema interpretation follows the client context, allowing clients that reject an explicit draft-07 declaration to consume the tool surface.

Domains:

• Workspace: create, inspect, update, delete, and traverse workspaces

• Organization: collections, collection-rule sync, workspace blueprints, and experimental organize or folder helpers

• Documents: search, read, create, publish, move, tag, custom properties, import/export, semantic composition, template inspection and native instantiation, capability and fidelity reporting, and block-level mutation

• Databases: create columns, add rows, update rows, inspect schema, add table/kanban views, and compose database structures from intent

• Comments: list, create, update, delete, and resolve

• History: version history listing

• Users and authentication: current user, sign-in, and profile/settings

• Notifications: list and mark notifications as read

• Blob storage: upload, delete, and cleanup blobs

For new document content, use create_doc for an optional single plain-text paragraph and create_doc_from_markdown for native headings, lists, links, code blocks, and $$...$$/$...$ math equations. Both tools accept folderId for immediate organize-folder placement. New pages record the authenticated AFFiNE account in the app's Created by property, including semantic pages, template instances, and workspace welcome pages. Existing creator records are preserved; older pages are not backfilled. New database rows also record the authenticated AFFiNE account and creation time, including rows composed from intent or inserted as paragraph blocks. Editing or moving existing rows preserves their creation metadata; older rows are not backfilled.

Use AFFINE_TOOL_PROFILE=read_only, core, or authoring when a deployment should expose a smaller surface than the complete full default. This is the recommended path for hosted, browser-connected, or least-privilege deployments because it reduces agent choice overload while keeping the full tool catalog available as an opt-in surface. You can also combine profiles with AFFINE_DISABLED_GROUPS such as docs.database, destructive, or admin for finer control.

Full-note replacement with replace_doc_with_markdown is destructive and requires full without disabling the destructive group. core and authoring retain incremental editing through append_markdown and update_block.

For the grouped catalog, notes, and operational caveats, see docs/tool-reference.md.

Documentation Map

Verify Your Setup

Useful CLI commands:

• affine-mcp status - test the effective configuration

• affine-mcp status --json - machine-readable status output

• affine-mcp doctor - diagnose config and connectivity issues

• affine-mcp workspaces [--json] - list accessible workspaces by name and mark the default

• affine-mcp workspace [id] - validate and set the local default workspace

• affine-mcp show-config - print the effective config with secrets redacted

• affine-mcp config-path - print the config file path

• affine-mcp snippet <claude|cursor|codex|all> [--env] - generate ready-to-paste client config

• affine-mcp logout - remove stored credentials

Core configuration uses environment > saved config > defaults. HTTP body/session limits, proxy settings, WebSocket timeouts, and the HTTP compatibility escape hatches are environment-only runtime flags; they are not read from the saved KEY=value file. See configuration precedence and environment scope. For a self-hosted deployment with a non-standard GraphQL route, use affine-mcp login --graphql-path /your/graphql/path or set AFFINE_GRAPHQL_PATH; show-config --json prints the exact resolved graphqlEndpoint without exposing secrets. Generated workspace and document URLs use the configured AFFiNE base URL, so a custom GraphQL path does not become part of a browser link.

For common failures, see:

• docs/getting-started.md#common-first-run-failures

• docs/configuration-and-deployment.md#deployment-checklist

Security and Scope

• Never commit passwords, session cookies, or compatible bearer tokens

• Use a dedicated least-privilege AFFiNE account for unattended deployments

• Email/password HTTP sessions share one login and never fall back to anonymous backend requests after authentication failure

• Use HTTPS for non-local deployments

• Keep remote HTTP MCP listeners authenticated; bearer mode refuses a non-loopback bind without AFFINE_MCP_HTTP_TOKEN

• Send MCP bearer tokens in the Authorization header, never in the URL

• Re-run affine-mcp login when a saved browser session expires

• For GUI clients, verify command -v affine-mcp and command -v node; an absolute script path does not fix a #!/usr/bin/env node shebang when the GUI process has no Node.js directory in PATH. See GUI client PATH troubleshooting.

• Restrict exposed tools with AFFINE_DISABLED_GROUPS and AFFINE_DISABLED_TOOLS for least-privilege setups

• Treat OAuth mode as a shared AFFiNE service-account deployment: it defaults to read_only, and write-capable profiles require AFFINE_OAUTH_ALLOW_SERVICE_WRITES=true

• Use /healthz and /readyz when running the HTTP server behind a container platform or load balancer

• Set HTTP body, session, idle, and shutdown limits explicitly for high-volume deployments

Development

Run the main quality gates before opening a PR:

npm run ci

Additional validation:

• npm test verifies tool metadata, test-suite coverage, and the fast regression suite without requiring a live AFFiNE instance

• npm run test:comprehensive boots a local Docker AFFiNE stack and validates the tool surface

• npm run test:e2e runs Docker, MCP, and Playwright together

• npm run test:playwright runs the Playwright suite only

• Focused runners for the new high-level tool surface include npm run test:create-placement, npm run test:capabilities-fidelity, npm run test:native-template, npm run test:mutation-ack, node tests/test-database-intent.mjs, node tests/test-semantic-page-composer.mjs, node tests/test-structured-receipts.mjs, node tests/test-organize-tools.mjs, and node tests/test-supporting-tools.mjs

Live tests can mutate or delete AFFiNE data. They allow loopback targets by default and refuse non-loopback targets unless the disposable target is explicitly enabled and confirmed as documented in CONTRIBUTING.md. Never run them against production.

Local clone flow:

git clone https://github.com/dawncr0w/affine-mcp-server.git
cd affine-mcp-server
npm install
npm run build
node dist/index.js

Release Notes

• CHANGELOG.md

• RELEASE_NOTES.md

• GitHub Releases

License

MIT License - see LICENSE.

Support

• Open an issue on GitHub

• Review AFFiNE product documentation at docs.affine.pro

Acknowledgments

• Built for the AFFiNE knowledge base platform

• Uses the Model Context Protocol specification

• Powered by @modelcontextprotocol/sdk