Skip to main content
Workspace admins connect credentials; every run in the workspace uses them. Evaluations need a model credential and a sandbox credential. Claude and Meta sign-ins can’t generate tasks, so connect a model API key or a ChatGPT sign-in for generation. Every model key and sign-in reaches its model through SelfBench’s gateway, so the credential never enters the agent’s sandbox.

Custom Routers

Use a Custom Endpoint credential with the router’s HTTPS base URL, not a method’s URL. The hostname must be in the deployment’s SELFBENCH_CUSTOM_MODEL_HOSTS. Each harness calls the endpoint in its own protocol: Pi uses Chat Completions, Codex uses Responses, and Claude Code uses Messages. Give the base URL without /v1 for Claude Code, and without /chat/completions or /responses for any harness. A call is priced from what the router reports. If it reports none, it is priced at the list rate of the model its response names. A call whose model the catalog doesn’t price leaves its trial without a cost. To evaluate Jev through Dari, create a Dari router with Jev as its selector and add its base URL as a separate credential. The key and model ID stay the same.