Connect Your Existing AI Stack

Keep the models and runtime you already operate. An AI Power Grid worker runs beside that service, receives compatible jobs, and returns the result. It does not replace your inference engine or upload your model files.

Grid applications
       |
AI Power Grid Core
       |
outbound worker connection
       |
Grid worker sidecar
       |
your Ollama / vLLM / SGLang / ComfyUI runtime
       |
your GPU

Current Compatibility

Open means the current public text worker can connect. Qualification means owning the runtime is not enough: the exact model, recipe, dependencies, and local canary must pass the media release process before the capability may be advertised.

What You Control

For the public text worker, the operator chooses:

• which backend endpoint the worker may call;
• which local model is advertised to the Grid;
• the maximum output and context length;
• maximum simultaneous Grid jobs;
• optional local-time operating windows, including paused windows;
• when the worker starts, drains, stops, or reconnects.

The media manager exposes the same basic availability decision with a smaller, honest control surface: always available, paused, or paused during one bounded maintenance window. Media execution remains one job at a time. Changing the availability policy updates a running worker within a few seconds; pausing stops new claims and lets the active render finish.

The worker uses an outbound WebSocket connection to receive jobs. Backend URLs default to loopback but may point at an operator-controlled LAN or remote API. Remote API credentials remain on the worker host.

Security Boundary

• Model files remain inside the operator’s runtime.
• The worker needs a worker-scoped Grid credential, never a wallet private key.
• The payout wallet is configured or delegated separately through Console.
• Discovered models are not authority to invent unsupported Grid model names.
• Media workers receive only short-lived upload URLs, not Grid storage keys.

Existing Stack Path

1. Start the backend you already use.
2. Download the verified worker from aipowergrid.io/run.
3. Let the local wizard detect the endpoint, or enter it manually.
4. Choose the exact model and limits you want to advertise.
5. Run the local backend test.
6. Approve the worker-scoped Grid connection in Console.
7. Choose the availability and concurrency you are willing to provide.
8. Wait for the local dashboard to report Online.
9. Run the parameter-free Grid test. Core chooses the payload and routes it only to that exact worker; the result is connectivity evidence, not a model quality score, charge, reward, or validator vote.
10. Confirm the worker by exact name or ID using the public check on /run.

Starting Fresh

Use the browser-only hardware planner at aipowergrid.io/run for an initial path. The downloaded worker remains authoritative for local compatibility. A browser guess cannot approve hardware, advertise a model, or qualify a media profile.

Backend Guides

Ollama

Start Ollama and make sure ollama list shows the model you intend to serve. The worker detects the default loopback endpoint and lists available models in its setup wizard. Install Ollama only from its official download page.

Open the Ollama quickstart.

vLLM, SGLang, and LMDeploy

Start the runtime’s OpenAI-compatible server, then enter its base endpoint in the worker. The worker probes the models endpoint before saving it. If the backend requires authentication, enter that backend credential only in the local worker UI.

Open the vLLM quickstart or SGLang quickstart. LMDeploy operators can use the LMDeploy quickstart.

LM Studio

Enable LM Studio’s local server, load a model, and leave the server running. The worker normally detects the default local endpoint; otherwise enter the endpoint shown by LM Studio.

Open the LM Studio quickstart.

KoboldCpp

Start KoboldCpp with its API enabled. The worker probes its model endpoint and bridges compatible text jobs through the same local setup flow.

Open the KoboldCpp quickstart.

ComfyUI

An existing ComfyUI installation is useful, but it does not make every model or workflow Grid-compatible. The bridge must resolve an approved model name to the exact checkpoint, nodes, and workflow recipe required by Core. Follow the media qualification guide rather than advertising arbitrary custom workflows. Media workers currently serve one job at a time and support local-time windows that pause new Grid claims without interrupting the active render.

Open the ComfyUI qualification guide.

ACE-Step

ACE-Step uses the media manager’s pinned direct-runtime profile. That public profile remains qualification-gated until its required hardware evidence, recipe commitment, signed release, and supervised staging are complete.

Open the ACE-Step qualification guide.

Troubleshooting

The backend is not detected: confirm it is running, then enter its endpoint manually and use Test. Containerized workers may need the host gateway instead of 127.0.0.1.

The model is missing: confirm the backend itself lists or has loaded the model. The Grid worker does not make an unavailable model appear.

The worker runs but is not online: a running process is not proof of Grid registration. Check the dashboard connection state and redacted logs, then use the exact-name public worker check.

No jobs arrive: traffic and model demand vary. Confirm that the advertised Grid model name matches a currently supported route and that the operating schedule is not paused.