Deploy AgenticMind¶

Run AgenticMind in production from the published GHCR images — no clone, no build, one OpenAI key, one shared MCP key.

Images: ghcr.io/moai-team-llc/agenticmind-server and …-worker, tagged per release (:latest, :0.5, :0.5.0). Requires Docker Compose v2.23+.

A. Standalone stack¶

cd deploy
./gen-secrets.sh            # generates AGENTICMIND_DB_PASSWORD + MCP_API_KEY into .env
#  then set OPENAI_API_KEY in deploy/.env
docker compose up -d

The stack (deploy/docker-compose.yml):

Service	Role
`agenticmind-db`	Postgres 17 + pgvector/pgvectorscale (extensions auto-created)
`agenticmind-migrate`	applies Drizzle migrations once, then exits
`agenticmind-server`	the MCP endpoint on `http://localhost:3000/mcp`
`agenticmind-worker`	the daily compounding sweep

Point your app at the MCP endpoint:

POST http://localhost:3000/mcp
Authorization: Bearer <MCP_API_KEY>

Strict profile (least-privilege database role)¶

The base stack connects the server as the postgres superuser, which row-level security does not apply to. For multi-tenant use, add the strict overlay:

./gen-secrets.sh            # also writes AGENTICMIND_APP_DB_PASSWORD
docker compose -f docker-compose.yml -f docker-compose.strict.yml up -d

Migrations still run as the owner; a one-shot agenticmind-app-role service provisions the agenticmind_server LOGIN role (a member of agenticmind_app), and the server connects as it with AGENTICMIND_PROFILE=strict — it refuses to start if pointed at a superuser or a BYPASSRLS role. See the security model.

B. Embed in your own compose¶

Copy the agenticmind-* services plus the configs: and volumes: blocks from deploy/docker-compose.yml into your stack, and add the three secrets to your own .env. Your app's container then reaches the server at http://agenticmind-server:3000/mcp.

Auth — one shared key, no minting¶

MCP_API_KEY is the simple single-tenant auth: one long random bearer your app sends. No issue-token, no AUTH_SECRET, no DB token row. It grants knowledge:read/write/admin, memory:read/write/admin and audit:write — not knowledge:signal, hitl:request or hitl:respond. For least-privilege or multiple distinct clients, set AUTH_SECRET instead and mint scoped JWTs with issue-token (see the README).

LLM — reuse your OpenAI key¶

The compose points CHAT_BASE_URL at api.openai.com/v1 with your OPENAI_API_KEY; no OpenRouter key is needed. Embeddings run locally (zero-key). Behind a blocked Hugging Face CDN, see OPERATIONS.md § Switching model providers.

Secrets¶

./gen-secrets.sh generates AGENTICMIND_DB_PASSWORD (random) and MCP_API_KEY (random) into deploy/.env and is idempotent — re-running keeps existing values. You never pick a database password by hand; the only secret you supply is the OpenAI key you already have.