---

name: deployment-configuration description: "Guides SuperAGI Docker, local runtime, configuration, migrations, GUI proxy, and optional GPU/local LLM deployment tasks." disable-model-invocation: true metadata: disco-role: operating license: MIT

SuperAGI Deployment and Configuration

Use this sub-skill when the task asks how to install, configure, start, debug, or choose between SuperAGI runtime modes. Prefer static checks and config validation before starting services.

Read First

• references/configuration.md for config.yaml, environment overrides, credential placeholders, and service hostnames.
• references/docker-stack.md for Docker Compose service topology, default vs GPU deployment, startup order, and migrations.
• references/local-runtime.md for non-Docker launchers and why they are side-effectful.
• references/troubleshooting.md for runtime, Docker, DB/Redis, GUI, NLTK, and GPU/local LLM failures.
• scripts/check_compose_files.py to statically inspect a user's compose files without running containers.

Main Decisions

1. Cloud vs local: Cloud use only needs provider/API keys in the hosted UI. Local use needs a checkout, config, Docker or Python dependencies, database, Redis, backend, worker, GUI, and proxy.
2. Docker vs host-local: Prefer Docker Compose for ordinary local operation because the repo's Dockerfiles encode Python 3.10, system packages, NLTK downloads, entrypoint order, and service hostnames.
3. Default vs GPU Compose: Use the GPU compose path only when the user explicitly needs local LLM/GPU support and has NVIDIA Docker runtime ready. Ordinary OpenAI/hosted-provider workflows do not require CUDA.
4. Service startup authorization: Starting the stack can build images, download Python/npm packages and NLTK data, run migrations, create volumes, and start long-running services. Ask/confirm unless the downstream task clearly authorizes this.

Safe Workflow

1. Confirm a target checkout has config_template.yaml, docker-compose.yaml, main.py, and superagi/.
2. Create or inspect config.yaml. Run the root config checker before using real credentials or starting services.
3. Use the compose checker here to verify expected services and optional GPU reservations.
4. If starting the default stack is authorized, ensure Docker/Compose is available, then use the checkout's default compose file. Expect public GUI access through the proxy on port 3000 and backend app startup on port 8001 inside the stack.
5. If using GPU/local LLM, read the GPU notes in docker-stack.md first; do not treat visible host GPUs as proof that Docker GPU passthrough works.

Boundary Notes

• Endpoint semantics, auth, webhooks, and SQLAlchemy models belong to api-service.
• Agent execution, workflow seeds, prompt parsing, and Celery job behavior belong to agents-workflows.
• Toolkit downloads and dependency installer behavior belong to toolkits-integrations.
• Provider keys, resources, and vector stores belong to models-resources-vector.