yuuno-mcp

MCP server for yuuno.tech adflow — drive the node-based AI media canvas from an MCP client (Claude Code, Claude Desktop, Cursor, …). You become an agent on the board: create sessions, wire up image/video/audio/text generation nodes, run flows, and read results — the same canvas the in-app chat edits, so your changes appear live in the browser.

Get an API key

Mint a workspace key at app.yuuno.tech → Settings → Organization → API (https://app.yuuno.tech/settings/organization/api). It's a 40-character hex string scoped to your workspace. Rotating it (Reset) invalidates the previous key for every connected client in that workspace.

Connect

Claude Code

Local (stdio, recommended — works everywhere):

claude mcp add yuuno -e YUUNO_API_KEY=<your-40-hex-key> -- uvx yuuno-mcp

Hosted (Streamable HTTP — no local install):

claude mcp add --transport http yuuno https://mcp.yuuno.tech/mcp \
  --header "Authorization: Bearer <your-40-hex-key>"

Claude Desktop

Add to claude_desktop_config.json (stdio; Desktop connectors don't pass custom headers, so use stdio):

{
  "mcpServers": {
    "yuuno": {
      "command": "uvx",
      "args": ["yuuno-mcp"],
      "env": { "YUUNO_API_KEY": "<your-40-hex-key>" }
    }
  }
}

Cursor

Add to ~/.cursor/mcp.json (or a project .cursor/mcp.json):

{
  "mcpServers": {
    "yuuno": {
      "url": "https://mcp.yuuno.tech/mcp",
      "headers": { "Authorization": "Bearer ${env:YUUNO_API_KEY}" }
    }
  }
}

Cursor also supports stdio ("command": "uvx", "args": ["yuuno-mcp"]).

Client support note: remote (hosted HTTP + Authorization header) works today in Claude Code and Cursor. Claude Desktop and claude.ai custom connectors don't pass static headers — use stdio (uvx) for Desktop; OAuth for claude.ai is a later release.

First moves for the agent

Call get_cheatsheet once at the start of a session — it explains the 18 node types, the DAG op rules, the run lifecycle, and model selection. Then a typical build is: list_models → create_session → apply_ops_batch (build the graph) → auto_layout → estimate_credits → run_flow → wait_for_run. list_examples returns ready-to-apply op recipes.

Tools (37)

• Sessions — whoami, create_session, list_sessions, get_dag, copy_session, delete_session
• DAG ops — add_node, add_node_with_inputs, update_node, remove_node, add_edge, update_edge, remove_edge, apply_ops_batch, set_uploaded_assets
• Structure — add_group, move_to_group, ungroup, replace_edge, insert_node_on_edge, bypass_node, list_node_io, auto_layout
• Runs — run_flow, get_run_status, list_runs, wait_for_run, estimate_credits
• Generate / catalog — execute_utility_node, wait_for_task, generate_text, generate_asset, list_models, get_model_schema, upload_asset
• Docs — get_cheatsheet, list_examples (also exposed as MCP resources yuuno://cheatsheet and yuuno://examples/<name>)

Configuration

In hosted HTTP mode the key is never held server-side — each request carries its own Authorization: Bearer <key>, verified against the backend, and the server proxies that exact key to Django (no cross-tenant state).

Development

cd mcp
python3.12 -m venv .venv
.venv/bin/pip install -e '.[dev]'
.venv/bin/python -m pytest tests -q

The contract-parity test (tests/test_parity.py) imports adflow.dag_protocol.constants from the monorepo root and asserts the mirrored constants match; it auto-skips outside the monorepo. Deploy artifacts (systemd units, nginx config, operator runbook) are under deploy/.