Connect KoboldCpp

The text worker can bridge a KoboldCpp server you already run. Its common local endpoint is http://127.0.0.1:5001; the worker detects the loaded model through /api/v1/model and uses the compatible generation API.

Install or Start KoboldCpp

Keep your existing KoboldCpp binary and launch settings. New operators should download a release from the official KoboldCpp repository, load a model, and enable its API before starting the Grid sidecar. The worker does not download or manage the model.

Detect the Endpoint

The common endpoint is http://127.0.0.1:5001. The worker reads /api/v1/model; enter a different operator-controlled base URL manually when you changed the port or host.

Compatible Models

Select the exact loaded model reported by KoboldCpp. The worker requires a successful local generation test; endpoint detection alone does not prove that the compatible generation path returns useful output.

Connect and Configure

1. Start KoboldCpp with its API enabled and a model loaded.
2. Download the verified text worker from aipowergrid.io/run.
3. Select detected KoboldCpp, or enter its operator-controlled base URL.
4. Choose the model reported by the backend and run the local generation test.
5. Set conservative context, output, schedule, and concurrency limits.
6. Approve the worker-scoped connection in Console.
7. Start the worker, wait for Online, and verify its exact name on /run.

Operator Controls

Configure the selected model’s context/output limits, maximum Grid concurrency, and optional local-time schedule. Start with one Grid job. Pause windows stop new claims and allow active work to drain; they do not monitor unrelated local requests automatically.

Security Boundary

Model files and any backend credentials remain local. The management endpoint defaults to loopback and the worker connects outbound to the Grid. Do not put Grid keys, backend keys, or wallet secrets in command arguments or public support logs. Community workers receive plaintext prompts and outputs.

Verify on the Grid

Wait for Online, then use the exact-name or worker-ID check on aipowergrid.io/run. Confirm that Core reports the same model you selected locally.

Troubleshooting

If detection succeeds but generation fails, test the backend from its own UI and confirm its compatible API is enabled. Detection proves endpoint inventory, not that a full generation will complete.

• Not detected: confirm the API is enabled and the base URL reaches /api/v1/model.
• Wrong model shown: unload the stale model, load the intended one, and re-scan before saving setup.
• Local clients slow down: lower Grid concurrency or add pause windows.