Hosted & Custom Models
Add accounts under Settings → Models & Providers.
ChatGPT-plan access
Choose ChatGPT, install the optional Codex component if prompted, and complete OpenAI's managed browser sign-in. Locus discovers the plan's visible models and usage windows. This route does not use or fall back to an OpenAI API key.
Each account has a Codex-native mode toggle:
- New ChatGPT accounts begin with the Locus contract—Locus's prompt and tools, approved memory, cross-chat context, and skill index.
- Existing accounts keep the setting they had before the 2.1 update.
- Turn Codex-native mode on for the model's Codex prompt, voice, and Codex-shaped shell, patch, and plan tools. Locus still owns execution, permission prompts, deny lists, and edit previews.
- Codex-native chats deliberately omit approved memory, cross-chat context, and the skill index.
- Changing the toggle restarts that conversation's server-side context. The first message in an existing ChatGPT chat after the update may replay its history once.
The account editor also controls model-supported reasoning effort and optional OpenAI web search. Reasoning effort applies on the next turn without resetting the conversation; higher effort can use plan quota faster. Web search is off by default and sends search queries to OpenAI when enabled.
Ask, explicit teams, evaluations, and non-ChatGPT providers keep their existing behavior.
Claude plan preview
Choose Claude plan under Settings → Models & Providers to use the managed Claude subscription route. Claude API remains a separate account type with a separate API key and billing. Existing API and ChatGPT accounts keep their settings.
Install the verified Claude component if prompted, then complete its managed sign-in flow. Each account has a separate credential home; Locus does not import the credentials from an existing Claude installation. Sign-in, sign-out, and refresh belong to the managed helper. Remote accounts sign in independently on the destination host.
Claude plan conversations can resume their managed session between turns. Locus supplies its tools and retains permission checks, mode restrictions, team roles, and saved-agent access ceilings. Forked conversations rebuild from their copied transcript rather than sharing a live provider session. Saved-agent instructions are sent through the system channel and are not duplicated in replayed user messages.
Available model aliases and reasoning choices come from the helper. Subscription usage is recorded when exposed; missing quota information is unknown, not zero. API-equivalent estimates are not treated as subscription charges. Claude plan does not itself provide image generation; choose a separate image account.
Preview limitation: distributed subscription-login approval has not been confirmed. An authenticated saved-agent greeting has been exercised, but clean-machine component installation, full login/logout/expiry, concurrent accounts, attachments, tools, recovery, scheduling, and remote sign-in still have open verification gates. Treat the route as a preview, not a claim that these cases are production-verified.
An incomplete managed response can leave an uncertain outcome. Inspect resulting files and actions before continuing in a new or forked task; retries cannot silently replay uncertain work. A remote runtime without the optional Claude helper does not advertise Claude plan support.
API-backed accounts
| Account | Endpoint | Authentication |
|---|---|---|
| OpenAI API | api.openai.com/v1 | OpenAI API key |
| Claude API | api.anthropic.com/v1 | Anthropic API key |
| Kimi | api.moonshot.ai/v1 | Moonshot API key |
| Kimi Code | api.kimi.com/coding/v1 | Kimi Code subscription key |
| Custom | Your URL | Endpoint key if required |
Use Test Connection before selecting the route. Locus reports a rejected key, an incorrect URL, or a waking endpoint. When a provider has no model-list route, the test sends a minimal completion.
Multiple accounts for one provider remain distinct. Model and account changes requested during a run wait until it finishes. Removing the active hosted account returns the session to local Ollama. In 2.1, switching models no longer checks hosted ChatGPT or Kimi models against the locally installed Ollama list.
Custom endpoints and safety
Custom accounts support vLLM, TGI, llama.cpp, Hugging Face endpoints, and compatible services. Authenticated non-loopback endpoints must use HTTPS. The 3.0 transport configuration preserves access to configured self-hosted model servers on private networks without a blanket transport-security exception. Redirects are refused so credentials cannot follow a response to another origin.
If the endpoint rejects tool calling, Locus retries once without tools and says the route can answer but cannot edit. If it rejects an image, Locus retries once without images and keeps that route text-only for the session.
Selecting a hosted route sends the prompt and included project context to that provider. Automatic hosted team routing requires separate one-time consent.