Everything configurable lives on the Settings page — providers, keys, and models.
Go to Settings → 模型供应商 and click add. Each provider needs three things:
https://api.deepseek.com/v1/chat/completionsdeepseek-chatOnly http(s) endpoints are accepted — the app refuses file:// and other schemes.
When you reopen Settings, existing keys appear masked (sk-1…cdef). Leave the field untouched to keep the stored key; type a new value to replace it; clear it to delete. Saved providers persist in the database — they survive restarts and redeploys.
The test button pings the provider with a minimal request using its stored key and stored endpoint — for safety, a stored key can never be sent to a different URL you type in. To probe an arbitrary endpoint, paste the key explicitly.
The 参与模型 selector on the main page chooses how many providers join each review: the full pool, or a random subset of 2–8. Subsets are resampled per run — handy for cheap spot-checks against a large pool.
Two or three inexpensive models (e.g. DeepSeek + a small OpenRouter model + Groq) give you real cross-verification for pennies per review. Add a frontier model as the adjudicator when quality matters most.