AppFlowy MCP Server

An open-source Model Context Protocol server for AppFlowy Cloud — so AI agents can read, build, and finish work in your AppFlowy workspace.

Not just view a Kanban board: create databases and pages, move cards, edit and delete rows, and edit document blocks (including advanced ones like callouts and toggles) — all over a single endpoint that's scoped to exactly the workspaces you choose and gated by Google OAuth or a token.

ℹ️ Independent community project — not affiliated with or endorsed by AppFlowy. It talks to the public AppFlowy Cloud API with an account you control.

✨ Highlights

• Full CRUD, not read-only. Create pages/databases/fields/rows, and — via a collaborative-editing (CRDT) layer — edit, move, and delete any existing row or document block, including ones made by hand in the UI.
• Advanced blocks the REST API can't make. Callouts, toggles, quotes, code, columns, and more, placed programmatically.
• Scoped by design. ALLOWED_WORKSPACE_IDS makes every tool refuse anything outside your allow-list — even if the account can see more.
• Three ways to connect. Google OAuth sign-in, a Bearer header token, or a token-in-URL link for UIs that can't set headers. Works with Claude, Cursor, Antigravity, and any MCP client.
• Agent-ready. Ships an AppFlowy "knowledge pack" as MCP server instructions and a fetchable appflowy://guide resource (see KNOWLEDGE.md) so agents build clean, well-formed structures.
• Batteries included. stdio / Streamable-HTTP / SSE transports, Docker, tests, and CI.

🧰 Tools

Read — get_workspaces, get_workspace_folder, list_databases, get_database_fields, get_database_row_ids, get_database_row_details, get_page, get_page_markdown (page/card body as Markdown), list_updated_rows

Create & structure (plain JSON, no CRDT) — create_page, create_database (grid/board/calendar), create_database_view, add_database_field, create_database_row, upsert_database_row, append_blocks, create_space, move_page, duplicate_page, rename_page, trash_page, restore_page

Edit & delete — any row or block (collab/CRDT layer) — update_row_cells (e.g. move a Board card by setting its Status), delete_row, add_block (incl. advanced: callout / toggle / quote / code…), edit_block_text (renders inline Markdown), replace_text (find-and-replace, no block id), delete_block

Schema editing (collab/CRDT) — update_database_field, delete_database_field, add_select_option, delete_select_option, set_group_by (Board grouping)

Pages and row bodies both accept Markdown, rendered into real blocks (headings, nested lists, interactive checkboxes, tables, code, math) — create_page(markdown=…), append_blocks(markdown=…), or a row's document. Field-type ids, the block-tree schema, and the full block palette are documented in KNOWLEDGE.md.

📦 What you can build

AppFlowy's editor and API offer a large palette of blocks, views, and inline styles. Here's what this MCP creates today — the rest is on the roadmap below.

Legend: ✅ full · 🟡 partial · 🗺️ roadmap · ⛔ not an MCP capability

Block-by-block detail lives in KNOWLEDGE.md (§9 Coverage).

🏗️ How it works

  MCP client                      this server                    AppFlowy Cloud
 (Claude, Cursor,   ── OAuth / ──►  (you host it)   ── bot ──►   (your workspace)
  Antigravity …)      token                           account
                                   ├─ REST API .......... reads, create page/db/field/row, append blocks, markdown docs
                                   └─ collab/CRDT (pycrdt)  edit/move/delete any row or block, advanced blocks

The server logs into AppFlowy Cloud with a dedicated bot account and exposes MCP tools. Most operations use the JSON REST API; surgical edits/deletes and advanced blocks go through AppFlowy's collaborative (Yjs/CRDT) layer via a merging update, so they're safe alongside live editing.

🚀 Quick start

Local (stdio) — most private, nothing exposed

cp .env.example .env      # fill in APPFLOWY_EMAIL / APPFLOWY_PASSWORD
pip install -r requirements.txt
python server.py

Docker / VPS — for remote & unattended agents

cp .env.example .env      # fill in credentials + MCP_SECRET_TOKEN (and OAuth, optional)
docker compose up -d --build

docker-compose.yaml includes optional Traefik labels. Keep the port off the open internet — front it with an identity gate (e.g. Cloudflare Tunnel + Access) and/or the built-in auth. A guided deploy_vps.sh is included for a first deploy.

⚙️ Configuration (.env)

🔌 Connecting a client

Google OAuth (recommended for remote/shared use — no token in the link)

1. Create a Google OAuth Web client; set the redirect URI to https://YOUR_HOST/auth/google/callback.
2. Set GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET, ALLOWED_EMAILS, OAUTH_ISSUER and restart.
3. In your MCP client, add a custom connector with just the URL https://YOUR_HOST/mcp/ — it will trigger a Google sign-in.

Bearer header (technical clients / CLIs)

{ "mcpServers": { "appflowy": {
  "url": "https://mcp.example.com/mcp/",
  "headers": { "Authorization": "Bearer YOUR_MCP_SECRET_TOKEN" }
}}}

For stdio-only clients, bridge with npx mcp-remote https://mcp.example.com/mcp/ --header "Authorization: Bearer YOUR_MCP_SECRET_TOKEN".

Link (token in URL) (for UIs that can't set a header)

https://mcp.example.com/mcp/?token=YOUR_MCP_SECRET_TOKEN

Convenient, but the token rides in the URL/logs — treat the whole link like a password.

Local (stdio)

{ "mcpServers": { "appflowy": { "command": "python", "args": ["/absolute/path/to/server.py"] } } }

🔐 Security

AppFlowy Cloud auth is full-account (there are no scoped API keys), so:

• Use a dedicated bot account invited only to the workspace(s) you expose.
• Set ALLOWED_WORKSPACE_IDS — enforced on every tool.
• Gate the endpoint with OAuth (email allow-list) or MCP_SECRET_TOKEN, and keep the port off the open internet.

Full details and vulnerability reporting: SECURITY.md.

🧪 Development

pip install -r requirements.txt
ruff format . && ruff check .
pytest -q

CI runs lint, format-check, and tests on every push/PR.

🩺 Troubleshooting

ValueError: Cannot decode update: while reading, an unexpected value was found — a pycrdt 0.14.1 regression fails to decode some AppFlowy database collabs, which breaks every collab-layer structural op (delete_row, field/option edits) on the affected database while row cell-writes keep working. requirements.txt pins pycrdt==0.13.0 to avoid it — don't unpin past 0.13 until it's fixed upstream.

A write looks like it didn't apply — get_database_row_details reads a materialized view that can lag the live collab by minutes; the write is usually fine. update_row_cells confirms its own write against the collab (with retry) before returning, so its success result is trustworthy.

🧭 How this compares

LucasXu0/appflowy_mcp covers pages/folders/trash/favorites. This server adds database write-back, full row/block edit & delete via the collab layer, workspace scoping, and OAuth — aimed at agents that finish work on a board, not just read it.

🗺️ Roadmap

Content & blocks

• Richer editing — insert-at-position and search-and-replace by content (no block ids), plus rich inline formatting (bold / links / color) on in-place edits.
• Person mentions — AppFlowy's editor has no person-mention type yet (page mentions ship via [](mention:id)); revisit when self-hosting adds agents as workspace members.
• More blocks — columns, toggle headings, table of contents, block equation, Mermaid, link-to-page.
• Import files — PDF / Word / PowerPoint / Excel / audio → Markdown → AppFlowy blocks (via markitdown), reusing the Markdown converter.
• Media — file / PDF / image / video / audio upload, web bookmark, embeds.

Databases

• More views & layouts — List, Gallery, Chart, Feed; inline and linked database views.
• Filters, sorts, and field reorder on views.
• Pagination + filtering on large database reads.

🤝 Contributing

Contributions welcome — see CONTRIBUTING.md and CODE_OF_CONDUCT.md.

📄 License

MIT.