@yuuno-io/mcp

MCP (Model Context Protocol) server that lets Claude Code — or any MCP client — build, run, and inspect Adflow visual-flow canvases on a Yuuno backend.

You write, Claude builds. Tell it:

"Build a 3-step ad pipeline: a Flux 2 pro image of a runner mid-stride → Kling 3.0 turning it into a 5-second slow-mo → Remotion videoEditor for the final cut. Save as 'Runner ad'."

…and the canvas appears in your Yuuno SPA in real time.

Install

# Run on demand (no install)
npx -y @yuuno-io/mcp

# Or install globally
npm install -g @yuuno-io/mcp
yuuno-mcp

Configure Claude Code

Add to your Claude Code MCP config (~/.config/claude-code/mcp.json or project-local .claude/mcp.json):

{
  "mcpServers": {
    "adflow": {
      "command": "npx",
      "args": ["-y", "@yuuno-io/mcp"],
      "env": {
        "YUUNO_BACKEND_URL": "https://api.yuuno.app",
        "YUUNO_API_KEY":     "yk_…paste your key here…"
      }
    }
  }
}

Restart Claude Code; you'll see adflow in the MCP servers list with 15 tools available.

Get an API key

1. Sign in to your Yuuno account.
2. Go to Settings → API Keys, click New key, give it a name (e.g. "Claude Code MCP").
3. Copy the yk_… value — the server shows it once.
4. Paste into your MCP config above.

If the SPA is unavailable, you can mint one over HTTP:

curl -X POST https://api.yuuno.app/users/me/api-keys/ \
  -H "Authorization: Bearer $YOUR_JWT" \
  -H "Content-Type: application/json" \
  -d '{"name": "Claude Code MCP"}'

The JSON response contains a key field — that's your YUUNO_API_KEY.

What the MCP can do

15 tools, all routed through the same apply_ops your SPA uses, so changes appear in the canvas immediately.

| Tool | Purpose | |---|---| | create_session(title) | New canvas | | list_sessions(limit) | Recent canvases | | get_dag(session_id) | Read full DAG | | delete_session(session_id) | Remove canvas | | add_node(session_id, node_type, …) | Insert one node | | update_node(session_id, node_id, patch) | Patch a node | | remove_node(session_id, node_id) | Drop a node | | add_edge(session_id, source, target) | Connect nodes | | remove_edge(session_id, edge_id) | Disconnect | | apply_ops_batch(session_id, ops) | Atomic multi-op | | list_node_types() | Catalog (12 node kinds) | | list_models() | Generation models catalog | | run_dag(session_id) | Execute the DAG | | get_run_status(run_id) | Poll execution | | improve_prompt(prompt, node_type) | Yuuno copilot helper |

All write tools accept an optional expected_version argument for optimistic locking — pass the save_version you read with get_dag and the server will reject stale writes with a 409.

Multi-tab / multi-user safety

Every API key is scoped to one user. The Yuuno backend filters every read and write by user_id, so your MCP cannot see or mutate sessions that belong to a different account. Concurrent edits from Claude Code + the SPA get serialized through SQLite + idempotent ops; the same op applied twice does not corrupt the canvas.

Domain portability

Nothing here hard-codes a domain. Point YUUNO_BACKEND_URL at any environment — local dev, staging, production — and the server keeps working. No rebuild necessary; the env vars are read at process start.

Development

git clone https://github.com/yuuno-io/yuuno
cd yuuno/mcp
npm install
npm run build           # produces dist/
npm run dev             # tsc --watch

Local backend (Yuuno API) is required for end-to-end testing — see the backend repo for spin-up instructions.

License

MIT — see LICENSE.