Provider connections

An administrator adds a connection under Connections: provider, label, base URL, API or internal engine key, and exact model IDs. API credentials are encrypted at rest and cannot be read back. Changing a connection key updates future requests without changing member keys. A connection's disabled state removes it from routing.

Choose Custom provider at the top of the provider menu to connect your own compatible endpoint. Give it a name, enter its API base URL and key, and choose Chat Completions, Responses or Messages to match the endpoint. You can discover models or enter their exact IDs. Existing compatible connections and their model IDs keep working.

Use Check & test models before adding or saving a connection. The check discovers the endpoint's model list and sends a small generation request for each model in each enabled request format. These requests consume provider quota and may incur charges. They run as administrator diagnostics and do not count toward a member's gateway usage or allowances.

Results show availability separately for each format, including denied access, missing models, unsupported test parameters, exhausted quota, rate limits and timeouts. A model is available when at least one format accepts a test; other formats can still fail. Only successful models are selected automatically. Review the selection and save the connection to make it available to members. Checks do not save credentials or configuration. Editing the key, endpoint or formats clears the previous result. A successful check confirms access at that moment, not future quota or every client feature.

Discovery follows native Anthropic and Gemini pagination and supports compatible data model lists. Gemini models that do not support content generation are shown as unsupported. If discovery returns 404, 405 or 501, manually entered IDs are tested instead. Discovery is bounded to 1,000 models and 30 pages, with an explicit incomplete-list notice. Tests use at most two concurrent requests, a 15-second timeout per request and a 15-minute limit per check. Stop check cancels remaining work; leaving the page also stops it. Credentials and raw provider responses are excluded from results. Results expire after ten minutes and are visible only to the administrator who started the check.

Official API connections

Discover models with the check button or enter exact IDs provided by your account. Use the per-format results to choose the formats your endpoint supports, then verify the intended client.

Claude uses the official API credential path. Anthropic documents organization gateways and restricts routing other users' requests through shared Pro/Max subscription credentials. Gateway documentation, credential rules.

GLM's general API and Coding Plan use distinct endpoints; the Coding Plan is documented for supported tools. Jatevo's default uses the general API. Z.AI endpoints. DeepSeek documents OpenAI-compatible access; this adapter implements its Chat path. DeepSeek API.

Optional bundled subscription engines

docker compose -f compose.yml -f compose.engines.yml up --build -d

Not run here: this block was not executed on the server that serves this page.

This runs pinned codex-lb and CLIProxyAPI privately. Enable subscription sign-in with ENGINES_ENABLED=1; existing personal accounts are not automatically registered as shared gateway connections. For a deliberate enrollment, use the operator CLI after connecting the accounts:

npm run cli -- accounts list
npm run cli -- connections enroll chatgpt --models '<model IDs>' --accounts '<explicit account IDs>'
npm run cli -- connections enroll grok --models '<Grok model IDs>'
npm run cli -- connections enroll kimi --models '<Kimi model IDs>'

Not run here: this block was not executed on the server that serves this page.

Run those commands with the private engine URLs and the same data directory. In Docker, use docker compose -f compose.yml -f compose.engines.yml exec console node cli.js ... instead of npm run cli -- .... ChatGPT enrollment creates an internal key pinned to the explicitly selected accounts. Grok/Kimi enrollment shares that provider's configured accounts on the private sidecar. Review the account inventory before enrollment; this is the deliberate transition from personal connections to team capacity.

For another private engine, use Connections or connections add. Engines receive internal keys, never the teammate's key. Keep a dedicated adapter endpoint per declared provider, particularly where multiple services expose the same model name.

Antigravity

The pinned CLIProxyAPI implements Antigravity OAuth and inference. Jatevo can route a connection to a separately configured Antigravity adapter using its internal API key. Authenticate and configure the adapter using the upstream project's instructions, then enter its private base URL and model IDs in Connections → Antigravity. Jatevo does not collect a Google password or expose a built-in Antigravity browser login.

Google's current consumer terms explicitly restrict third-party access through Antigravity login; the documented official path for external tools is a Gemini API credential. Antigravity's application currently documents no custom endpoint support. Technical adapter availability is not vendor endorsement. Google FAQ, plans, CLIProxyAPI.

Cursor

Cursor subscription inference uses a separate community adapter from the maintained CLIProxyAPIPlus fork; it is not in Jatevo's bundled original engine. Configure and authenticate that adapter independently, keep it private, then add Connections → Cursor with its internal key and exact model catalog. The Jatevo adapter handles gateway authorization and accounting around that endpoint. Subscription login and inference availability still depend on the external adapter and require live validation.

Inspected Cursor adapter source. Cursor's official APIs run agent workflows and administration, not a standalone model-inference API. Cursor APIs.

Cursor can separately consume Jatevo keys through BYOK. Calls transit Cursor servers, so the endpoint must be reachable with TLS. Cursor documents chat-only BYOK scope; Tab remains on Cursor's own models. Cursor BYOK.

Verification

CI validates routing, credential isolation and protocol usage with local mock endpoints. An integration is only live-verified for your deployment after actual authentication, a nonstreaming request, a streamed request, model/tool compatibility and usage reconciliation succeed with your credentials. Do not label a connection healthy solely because its configuration was saved.

This page as markdown · llms.txt