can i use custom provider like anything?

[mochadwi]

I do subscribe to commandcode and agnes.ai but don't see in the default/supported model and provider in here https://github.com/justrach/codegraff#providers--models is this something not possible on graff atm?

[justrach]

Yes — two ways today, depending on what "anything" is:

1. Any OpenAI-compatible gateway (broadest answer): set its key and go.
23 built-in providers are activated just by exporting their key — each has a fixed env_key in src/provider.zig:86. Notably the aggregator ones give you "anything":

• OPENROUTER_API_KEY → OpenRouter (hundreds of models, e.g. anthropic/claude-sonnet-4.6)
• AI_GATEWAY_API_KEY → Vercel AI Gateway (alibaba/qwen3.8-27b; coding-agent /v1)
• KILO_API_KEY, GROQ_API_KEY, CEREBRAS_API_KEY, DEEPSEEK_API_KEY, …

Run graff route to see everything reachable and what each bills; /model <name> or --model to pin.

2. Point a built-in provider at your own endpoint (self-hosted / proxied):

• GRAFF_XAI_URL — override xAI's base URL
• GRAFF_ZAI_URL / ZAI_CODING=1 — GLM coding-plan endpoint
• GRAFF_VERCEL_URL — Vercel gateway endpoint
• GRAFF_CODEX_URL — Codex-shaped mock/local endpoint

Wire kind matters for gateways: providers are .openai (chat completions), .anthropic (messages), or .responses — pick a provider whose wire your endpoint speaks.

What doesn't exist yet: a fully generic GRAFF_CUSTOM_URL + GRAFF_CUSTOM_KEY + GRAFF_CUSTOM_WIRE entry that invents a brand-new provider slot at runtime. If that's what you want, say so here and it's a reasonable small feature — a provider spec is one table row today.

[justrach]

Closing as answered, with one correction to the earlier reply: the generic mechanism already exists. .graff/.config.router (since ee5283a, documented in the README under the workspace-local router section) takes any base_url + env_key + default_model, so commandcode / agnes.ai work today if they expose an OpenAI-compatible /chat/completions endpoint — select it with graff --model <id>, pull its catalog with graff models refresh, set the key via graff key set or the env var. Current limits: one extra router per workspace, OpenAI chat wire only. Open a feature request if you need more than that.