Providers
9router-go supports 150+ providers and 1000+ models. Every provider is isolated and independent — no provider is merged, cross-aliased, or redirected into another.
The provider catalog
OAuth & CLI sessions (28)
Antigravity (antigravity), Antigravity CLI (ACP) (agy-acp), Claude Code (claude), Cline (cline), ClinePass (clinepass), CodeBuddy (codebuddy-intl), CodeBuddy CN (codebuddy-cn), Cursor IDE (cursor), Free Pi (free-pi), Freebuff (freebuff), GitHub Copilot (github), Grok CLI (Grok Build) (grok-cli), Kilo Code (kilocode), Kimi (kimi), kimi-coding (kimi-coding), MiniMax Code (minimax-code), MiniMax Code (Global) (minimax-code-global), Muse (Meta Model API) (muse), OpenAI Codex (codex), Qoder (qoder), Qoder CN (qoder-cn), trae (trae), windsurf (windsurf), xAI (Grok) (xai), Xiaomi MiMo (xiaomi-mimo), Zai GLM Coding (glm), ZCode (Z.ai) (zcode), Zed (zed)
Free tier (26)
Agnes AI (agnes), API.airforce (api-airforce), Bazaarlink (bazaarlink), BytePlus ModelArk (byteplus), Cloudflare (cloudflare-ai), DaoXE (daoxe), Devin CLI (devin-cli), Edge TTS (edge-tts), Gemini (gemini), Gemini CLI (gemini-cli), Google TTS (google-tts), Keenable (keenable), Kilo Gateway (kilo-gateway), Kimchi (kimchi), Kiro AI (kiro), Local Device (local-device), NaN.Builders (nan), Nous Research (nous), NVIDIA NIM (nvidia), Ollama Cloud (ollama), OpenAdapter (openadapter), OpenCode Free (opencode), OpenRouter (openrouter), Poolside (poolside), SearXNG (searxng), Vertex AI (vertex)
API key (92)
Alibaba (alicode), Alibaba Coding (alicode-intl), Alibaba Studio (alims-intl), Alibaba Token Plan (alitp-intl), Anthropic (anthropic), AssemblyAI (assemblyai), AtmoRouter (atmorouter), Atria Dawn (atria), AWS Bedrock (bedrock), AWS Bedrock (xAI) (bedrock-xai), AWS Polly (aws-polly), Azure OpenAI (azure), B.AI (bai), Baidu Qianfan (baidu), Black Forest Labs (black-forest-labs), Blackbox AI (blackbox), Brave Search (brave-search), Cerebras (cerebras), Chutes AI (chutes), Cohere (cohere), ComfyUI (comfyui), Command Code (commandcode), Command Code (Provider API) (commandcode-goat), Dahl Inference (dahl), Deepgram (deepgram), DeepRelay (deeprelay), DeepSeek (deepseek), ElevenLabs (elevenlabs), Exa (exa), Fal.ai (fal-ai), Featherless (featherless), Firecrawl (firecrawl), Fireworks AI (fireworks), Fish Audio (fish-audio), GLM (China) (glm-cn), Google PSE (google-pse), Groq (groq), HCNSec (hcnsec), HuggingFace (huggingface), Hyperbolic (hyperbolic), Inworld TTS (inworld), Jina AI (jina-ai), Jina Reader (jina-reader), Kenari (kenari), Linkup (linkup), LLM7 (llm7), Minimax (China) (minimax-cn), Minimax Coding (minimax), Mistral (mistral), Morph (morph), NanoBanana API (nanobanana), Nebius AI (nebius), Ollama Local (ollama-local), Ollama Search (ollama-search), OpenAI (openai), OpenCode Go (opencode-go), OpenCode Zen (opencode-zen), OpenDesign Cloud (opendesign), Perplexity (perplexity), Perplexity Agent (perplexity-agent), Recraft (recraft), RunCrate (runcrate), Runway ML (runwayml), SD WebUI (sdwebui), SearchAPI (searchapi), Self-Hosted Embedding (selfhosted-embedding), Self-Hosted STT (selfhosted-stt), Self-Hosted TTS (selfhosted-tts), SenseNova (sensenova), SenseNova (Intl) (sensenova-intl), Serper (serper), SiliconFlow (siliconflow), Stability AI (stability-ai), SumoPod (sumopod), Synthetic (synthetic), Tavily (tavily), Tencent Hunyuan (tencent), TinyFish (tinyfish), Together AI (together), Token Harbor (tokenharbor), TokenRouter (tokenrouter), Topaz (topaz), v1m (System One) (v1m), Venice AI (venice), Vercel AI Gateway (vercel-ai-gateway), Vertex Partner (vertex-partner), Volcengine Ark (volcengine-ark), Voyage AI (voyage-ai), Xiaomi MiMo (Token Plan) (xiaomi-tokenplan), Xquik (xquik), You.com Search (youcom), Z.ai Web Search (zai-search)
Web session cookie (4)
DeepSeek Web (deepseek-web), Gemini Web (gemini-web), Grok Web (Subscription) (grok-web), Perplexity Web (Pro/Max) (perplexity-web)
Connecting
Dashboard → Providers → pick a provider → Connect.
No signup needed:
- Kiro AI — around 50 free credits per month
- OpenCode Free — no auth
OAuth flows
Different providers authenticate differently, and the dashboard offers the right one per provider:
| Flow | Providers |
|---|---|
| Social / authorize-exchange | Kiro, Antigravity, Cline, Trae, Windsurf, Zed, Kimchi, Xiaomi MiMo |
| Device code | Any provider exposing a device flow |
| Local proxy callback | Codex |
| Bulk import | Codex, Grok CLI |
| API key paste | Kiro |
| Cookie / PAT paste | iFlow (cookie), GitLab (PAT) |
| Session polling | Freebuff — initiate, poll, inspect, and switch between sessions |
Built-in public client credentials exist for antigravity, gemini-cli, xai, codex,
iflow and others, so most operators never set an OAuth client ID. Override per
provider with {PROVIDER}_OAUTH_CLIENT_ID / {PROVIDER}_OAUTH_CLIENT_SECRET
when you run your own OAuth app.
Tokens are refreshed automatically, and a request that fails with 401 is retried
once after a refresh.
Browser-based login
Some web providers have no OAuth endpoint at all. Two paths:
- Gemini Web (
/api/providers/gemini-web/browser-login) and DeepSeek Web (/api/providers/deepseek-web/browser-login) open a browser, wait for the sign-in, and read the session out. - Import from a signed-in browser (
/api/providers/{gemini,deepseek}-web/import-chrome) reads the browser's own cookie store. Close the browser first — a browser that is still open rotates the same session from the other side and the two renewals cancel each other out, which signs the account out.
Renew a DeepSeek session with
POST /api/connections/{id}/renew-deepseek.
Custom providers
One shared dialog adds an OpenAI- or Anthropic-compatible endpoint. One custom
provider is one providerNodes row, and a node can have many connections; a
connection points at its node through provider = node.id.
| Field | Purpose |
|---|---|
| Name | Display name |
| Prefix | Wire syntax prefix (provider/model) |
| Base URL | Upstream endpoint |
| Headers | Your own request headers |
Validate an endpoint before saving with POST /api/providers/validate.
Per-provider limits
Each custom node can carry a capacity budget:
| # | Limit | Unit | Meaning |
|---|---|---|---|
| L1 | RPM | requests/minute | Request rate this endpoint tolerates |
| L2 | TPM | tokens/minute | Input (prompt) token rate this endpoint tolerates |
| L3 | Concurrency | parallel requests | Number of simultaneous requests |
| L4 | Lifetime quota | total tokens | Cumulative cap before the node is declared exhausted |
| L5 | Queue timeout | ms | How long a request may wait in the FIFO queue for L1/L3 |
| L6 | Contract expiry | RFC3339 UTC date | Validity period of the node's contract |
0 means unlimited; an empty string means no expiry — the same convention as
API keys.
Behaviour when exceeded: congestion (RPM/TPM/concurrency) marks the candidate
as full for that attempt only, and the router tries the next candidate; 429 is
returned only when every candidate is exhausted. Node state (quota exhausted,
contract ended) skips all of that node's connections and returns a typed 403.
A congestion burst writes no cooldown. One request over a limit must not
globally park a healthy node — only a debug trace is written. The lastError row
the dashboard's node-health table reads is still written, so you can see the
cause without the node being locked.
Set limits with PUT /api/provider-nodes/{id}/limits and lift a spent quota with
POST /api/provider-nodes/{id}/quota/reset (it moves the watermark, it does not
delete history).
Models
Browse a provider's models with GET /api/providers/{id}/models, export keys
with GET /api/providers/{id}/export-keys, and remove many at once with
DELETE /api/providers/{id}/models.
Custom models are stored in the customModels key-value scope, not as columns
on providerNodes. Add one at a time or paste many (separated by spaces, commas
or newlines) — one request, atomic, per-row results.
Aliases (/api/models/alias) let a client write a short name for a long
model id. Deprecations (/api/models/deprecations) surface models upstream
has retired, and auto-disable (/api/models/auto-disable) turns off the ones
that stopped working.
Important notes
- No cross-provider aliasing.
antigravityis notopencode; a model belonging to one provider is never redirected to another. - No substring-based model rewriting. Routing decisions are driven by catalog
registration, database aliases, custom node prefixes, or standard
provider/modelwire syntax — never by ad-hoc string inspection. - The Import from /models button only appears for providers that genuinely support dynamic catalog discovery upstream.