Third-party endpoints and keys in Orca, set per agent

Yes, because they never shared one config to begin with. Orca's supported page says the agent combobox just launches a process in a terminal, so the endpoint and the credential belong to whichever CLI got launched: Claude Code reads the env block in settings.json plus those 2 variables, Codex reads ~/.codex/config.toml, Droid reads its customModels entries in ~/.factory/settings.json, Qwen Code reads ~/.qwen/settings.json, and Cline and Roo Code read what you typed into their settings panels. On the Claude Code side the official line adds a rule: a gateway credential variable takes precedence over a saved claude.ai login or Console key, so your own key outranks the subscription. Cursor CLI is the counter-example: the docs give only browser login and an API key, and that key comes from Cursor's own dashboard. Checked 2026-09-22.

Because it never asks the vendor. The official usage-tracking page says Orca reads the local usage state each agent maintains on disk, followed immediately by No API calls, no extra auth., and the same page says numbers update when the agent writes, not in real time. The coverage that page names is 6 agents: Claude Code, Codex, Gemini, OpenCode, Kimi Code, MiniMax. Whether that readout stays correct after you point an endpoint at a third-party gateway is not written in the official docs, and this page does not guess for them. Look at the key provider's own side for the aggregated bill (checked 2026-09-22).

Because the docs give it to 3 agents only. The orchestration page says --model accepts opaque provider model IDs for Claude, Codex, and Cursor., and in the same sentence requires --effort to come with --model and to apply only when that agent and model support the level. Those commands also need Settings → Experimental switched on first. For the rest, a different endpoint or model is set in that agent's own config, which is what the 11 rows of the table on this page are for. Orca itself writes Command flags evolve with the app, so this shape holds only for the day it was checked, 2026-09-22.

Two layers. What Orca's docs commit to is about launch arguments and environment: new launches come with each supported CLI's permission-bypass flag pre-filled, and once you have overridden a specific agent the official sentence is Orca leaves that agent alone, followed by a rule that a non-empty custom value counts as an explicit override and opts that agent out of later permission-mode migrations. The docs never say whether some update will rewrite your agent config files, and there is no blanket guarantee to quote — those two sentences are all there is. What really drifts is field names and paths in vendor docs: the endpoint sits in your user-level file, and updating Orca is a different event from updating that CLI (checked 2026-09-22).

Start with the shape of base_url: Qwen's docs say that for a hosted OpenAI-compatible gateway you set baseUrl to the API's /v1 root, for example https://gateway.example.com/v1, not the full /v1/chat/completions path, since the SDK appends that path itself. Then check the family: Claude Code's Anthropic Messages format expects the endpoint to forward the anthropic-beta and anthropic-version headers unchanged. Then confirm the config was read: Grok's docs say to run grok inspect after editing ~/.grok/config.toml and see what it found there. Finally the credential slot: Claude Code's rule is ANTHROPIC_AUTH_TOKEN when the other side says bearer token or Authorization header and ANTHROPIC_API_KEY when it says API key or x-api-key, and when a shell export and the settings.json env block set the same variable, the file wins (checked 2026-09-22).

It differs per agent, and splits into files and settings panels. Claude Code: the official route is an env block in settings.json, plus an explicit warning not to put the credential in a project's .claude/settings.json, which gets committed to the repository. Codex: in the official schema env_key stores the name of an environment variable, not the key, and the inline field is marked discouraged for security reasons. Kilo: environment references resolve only in trusted locations, a project file committed to a repo cannot resolve them, and the docs end with Keep provider credentials in your global config. OpenCode: /connect writes credentials into ~/.local/share/opencode/auth.json. Cline and Roo Code: an API Key field in the settings panel. Grok also says Prefer env_key over hardcoding api_key. Orca stores none of this; it only launches that CLI (checked 2026-09-22).