---

name: platform-stack description: "Operate and troubleshoot the self-hosted AutoGPT Platform stack, Docker services, environment files, migrations, and local LLM setup." metadata: disco-role: operating disable-model-invocation: true license: NOASSERTION

Platform Stack

Use this sub-skill for AutoGPT Platform self-hosting, local development, Docker Compose, environment-file initialization, service health, migrations, ports, upgrades, and local OpenAI-compatible AutoPilot endpoints.

Start here

• Read workflows for the normal setup, separate backend/frontend development, migrations, upgrades, and local LLM routes.
• Read configuration before editing .env files, changing ports, or switching auth/model transport.
• Read troubleshooting before resetting data, diagnosing container restarts, or changing an existing installation.
• Run python scripts/platform_stack_preflight.py --repo <checkout> before starting services. Add --init-env only when missing local env files should be created without overwriting existing files.

Safe operating sequence

1. Confirm whether the task is a fresh self-host, active local development, an upgrade, or a production-like deployment. Do not mix these paths.
2. From the repository's autogpt_platform/ directory, verify Docker Compose, Node/Corepack, disk space, and the three expected default env files.
3. Initialize only missing .env files. Existing files must be diffed against current defaults; cp -n does not merge newly added variables.
4. Start the smallest service set that proves the next step. Use make start-core for Postgres, Redis, and RabbitMQ; use full Compose only when the UI or all services are required.
5. Apply migrations and generate Prisma artifacts before exercising backend features. Confirm docker compose ps and service logs before debugging UI symptoms.

Common commands

cd autogpt_platform
make init-env
make start-core
make logs-core
make migrate
make run-backend
make run-frontend
# Full stack, when explicitly wanted:
docker compose up -d --build
docker compose ps

The normal local entry points are frontend 3000, WebSocket 8001, and REST or execution API 8006, but the active Compose/environment configuration wins.

Routing boundaries

• Backend routes, blocks, Prisma, API schemas, and Python tests belong to platform-backend.
• Next.js pages, generated hooks, Builder/Copilot UI, and browser tests belong to platform-frontend.
• Classic autogpt, Forge, and direct benchmark commands belong to classic-agents.

Never make a Docker reset, database reset, credentialed provider call, model pull, or full browser suite a default smoke check. Ask for explicit scope when those side effects are part of the requested task.