Multi-agent assistant that coordinates schedule (MCP calendar), tasks (Google Tasks MCP), and notes (database + optional external-notes MCP).
| Path | Purpose |
|---|---|
| agents/loopie/ | ADK app: root_agent, tools, specialists |
| mcp_servers/ | MCP server (Google Calendar & Tasks; in-memory external notes) over SSE or stdio |
| web/ | Browser UI + API (uvicorn web.app:app); Docker: web/Dockerfile |
| sql/migrations/ | AlloyDB-compatible schema |
| deploy/CLOUD_RUN.md | Cloud Run, VPC, secrets, adk deploy, Loopie Web image |
Python 3.10+ is required (google-adk 1.19+ does not publish wheels for older versions). On macOS, prefer an explicit interpreter if python3 is still 3.9, for example python3.12 -m venv .venv.
See .env.example and deploy/CLOUD_RUN.md. For Vertex, set ADK_MODEL=gemini-2.5-flash and a supported region such as us-central1 if you see publisher model 404s.
python3.12 -m venv .venv && source .venv/bin/activate
python -m pip install -r requirements.txtUse python -m pip (not bare pip3) so installs always target the active venv.
Run MCP in terminal A:
PYTHONPATH=. python -m mcp_servers.app sseThen set MCP_SSE_URL, apply sql/migrations/001_init.sql, set DATABASE_URL, and start the Loopie browser UI/API:
uvicorn web.app:app --reloadDeploy only the agent folder:
adk deploy cloud_run --project=... --region=... agents/loopieThe app is single-tenant as shipped (one default user id and typically one Google OAuth token for MCP). Extending to many users requires per-user OAuth tokens and a partitioned or otherwise tenant-scoped notes database. See deploy/CLOUD_RUN.md §5.
The interpreter that ran pip is older than 3.10, or pip3 pointed at system Python instead of .venv (watch for “Defaulting to user installation”). Recreate the venv with Python 3.10+ and install with python -m pip install -r requirements.txt after source .venv/bin/activate. Check with python -V and which python.
MCP_SSE_URL must reach a running MCP SSE server. If nothing listens on that host/port, ADK fails when loading tools. Fix: start the MCP server from the repo root (PYTHONPATH=. MCP_PORT=8765 python -m mcp_servers.app sse) and keep the URL in sync, or set MCP_DISABLED=1 in .env to turn off MCP toolsets until the server is up.
Omitting MCP_SSE_URL also disables MCP; database tools still work when DATABASE_URL is set.
Set USER_TIMEZONE (IANA name, e.g. America/Los_Angeles) so the schedule specialist can resolve phrases like “next Wednesday” using real wall-clock context injected each turn.
