Supported model runners

Tokmine works with anything that speaks the OpenAI, Anthropic or Gemini API, or Ollama's native API.

Do not have a runner yet? tokmine install ollama (or llamacpp, via Docker) installs one and pulls a model. LM Studio is a desktop app you install yourself, see Install a local model.

How discovery works

1. On start, the CLI probes 127.0.0.1 on the well-known ports below, concurrently and with short timeouts.
2. For each server that answers, it probes the API dialects it speaks (OpenAI /v1/models, Ollama /api/tags, Anthropic, Gemini /v1beta/models). One server can speak several; Ollama, for example, speaks Ollama and OpenAI.
3. It lists each server's models and reports them, along with the dialects, to Tokmine.
4. It re-checks about every 30 seconds, so servers starting or stopping and models being pulled or removed are picked up without a restart.
5. Only servers it discovered or that you configured are ever proxied to. Tokmine cannot make the CLI reach arbitrary hosts.

Remote or custom servers: --provider

For a runner on another port or another machine (for example a GPU box on your LAN that cannot run the CLI itself), pass its base URL. The flag is repeatable.

tokmine --provider http://gpu-box:8000
tokmine --provider http://10.0.0.5:8080 --provider http://localhost:9000 --dialect openai
tokmine --no-discover --provider http://gpu-box:8000

The dialect is detected automatically. If detection guesses wrong, or the server hides its model list, force it with --dialect openai|anthropic|gemini|ollama. Add --no-discover to use only the servers you list. You can also set TOKMINE_PROVIDERS (comma separated) or store providers in the config file.

Checking what was found

tokmine models   # one-off scan, lists providers, dialects and models
tokmine status   # config, mode, endpoint reachability and providers

See Troubleshooting if a runner does not show up.