Skip to content

Providers and authentication

Supported backends include OpenAI-compatible providers, Anthropic, Gemini, llama.cpp, Ollama, and ChatGPT/Codex. Credential precedence is runtime input, then stored auth.json, then the provider environment variable. Common variables include OPENAI_API_KEY, ANTHROPIC_API_KEY, and GEMINI_API_KEY.

/thinking <off|on|minimal|low|medium|high|xhigh> and /thinking-cycle save the selected level as the default for future sessions; a loaded session continues to use its recorded level. Models without reasoning capability remain at off. off omits enabled reasoning where supported. OpenAI-compatible and Codex APIs use their available effort enums (their highest/lowest available effort may be used for xhigh/minimal); Anthropic and Gemini 2.5 use deterministic thinking-token budgets. Provider and model API restrictions still determine which controls a request accepts.

Temperature defaults to balanced (0.5). Presets are coder (0.2), balanced (0.5), creative (0.8), and experimental (1.2); numeric values are rounded to tenths and bounded to 0.0..1.2. The TUI shows the nearest mode and exact value and changes it with Ctrl+Left/Right or /temperature <value|mode>. Startup accepts --temperature and --temperature-mode; numeric wins when both are supplied. get_state exposes temperature through RPC, and set_temperature accepts a numeric temperature or preset mode. Subagents inherit their parent’s effective value unless their spawn_subagent call supplies temperature (which wins) or temperatureMode. Anthropic and Codex omit temperature and emit a warning; requests still run.

OpenRouter uses its native reasoning: {"effort": ...} request dialect and streams reasoning separately. Ollama Cloud uses native think; minimal and low map to "low", medium to "medium", high and xhigh to "high", and off to false. ChatGPT/Codex Responses reasoning summaries also render separately from answer text. Local Ollama and llama.cpp retain their OpenAI-compatible/local behavior.

For local models, no key is required:

In the TUI, press Ctrl+K, then choose Configure Ollama or Configure llama.cpp. Selecting an unconfigured local placeholder also opens setup. Enter the server URL, optionally click Fetch server models to list models already available on that server, then select a model or enter its exact ID manually. Save & use writes the model URL to global models.json, saves provider/model defaults in global settings.json, and activates the model. The dialog previews both paths and respects ONE_CODING_AGENT_DIR and XDG. To skip discovery, press Enter in the URL field to focus Manual model ID, type the exact model ID, and press Enter again to save. The form scrolls in compact terminals while Save/Cancel and progress/error feedback remain visible.

Run Ollama or llama-server with a loaded model separately; ONE does not install servers or download model weights. Opening the dialog makes no network request, and saving a manually entered ID does not verify server/model availability. From a container, use an address that reaches the host rather than container localhost; see Docker. Existing unrelated configuration is preserved; repair malformed files or conflicting project defaults before saving. The two files are written individually, not as a cross-file transaction; a partial save or activation failure is reported so you can review and retry. one doctor --setup remains an offline alternative for selecting defaults.

Terminal window
LLAMA_CPP_BASE_URL=http://127.0.0.1:8080 one --provider llama.cpp --model local
one --provider ollama --model local --ollama-url http://localhost:11434/v1

/login validates credentials before storing them and refreshes available models. Anthropic and ChatGPT/Codex also support subscription OAuth login; see Subscription login. Provider-controlled OAuth endpoints can change without notice.

  • /login anthropic subscription starts Anthropic OAuth. /login refresh anthropic refreshes models; /logout anthropic removes local stored credentials.
  • /login chatgpt subscription starts ChatGPT/Codex OAuth. Seed Codex models are available before a live model list is fetched; /login refresh chatgpt refreshes it.

/logout <provider> removes one-managed runtime/stored credentials locally. It does not revoke a provider account token or remove environment credentials.