Using the AI agent panel
Pick a model, describe the change, and review the agent's file diffs before anything is applied.
AI
Preview — early access. Provider coverage in the model router is still growing.
Overview
The agent panel on the right reads your open editor tabs as context, works from the instruction you give it, and proposes changes as a set of file diffs — it never writes to disk on its own. Every reply is tagged with the model that answered, e.g. [claude-sonnet-5 · auto].
How it works
A model picker at the top of the panel controls which model answers. Leave it on auto and the router picks for you: Haiku for trivial one-line fixes, Sonnet for standard single- or multi-file work, and Fable for genuinely hard, cross-cutting changes — based on your instruction's wording, how many files are in context, and how much context you've loaded. Or pick a specific model from the full router:
| Lane | Models | Notes |
|---|---|---|
| Anthropic ladder | Haiku 4.5, Sonnet 5, Opus 4.8, Fable 5 | Subscription-billed; powers the auto ladder end to end |
| BYOK cloud providers | Grok 4 (xAI), Kimi K2 (Moonshot), Gemini Flash / Flash-Lite (Google), Qwen via DashScope (Alibaba), GPT-4o mini (OpenAI), Codestral (Mistral), Command R (Cohere), DeepSeek Chat, plus Groq / Cerebras / Together / Fireworks / OpenRouter hosted lanes | Each needs its own provider key — add it in the Secrets vault |
| Local, free (Creytix Model Server) | Qwen2.5-Coder (3B / 7B / 32B), Llama 3.1 8B / 3.3 70B, Gemma2 (2B / 9B), DeepSeek-Coder-V2 16B, Phi-4, Mistral, and more | Runs on your own machine, keyless. The Creytix Model Server ships the engine and loads the weights we already have — nothing to install or pull |
The Models tab in the IDE shows this same list live, with a dot per model: ● a key is configured, ○ it needs one, ⬤ it's a local Creytix model, free and always available.
Steps
- •
Open the agent panel
Toggle "AI editor" in the IDE header if it isn't already open.
- •
Pick a model
Leave it on auto, or choose a specific model from the dropdown — check the Models tab first if you're not sure a provider's key is configured.
- •
Describe the change
Write your instruction. Reference open files by name, or send a picked element over from the Browser tab.
- •
Review the diffs
The agent proposes file-by-file changes — read them before applying.
- •
Apply, rate, or flag
Apply what you want, rate the response, or flag it if something looks off — feedback routes back into the platform's governance checks.
Capabilities
- Live model registry — the picker and the Models tab both read the same generated router registry, so what you see is what's actually wired.
- Auto model ladder — trivial, standard, and hard work route to Haiku, Sonnet, and Fable automatically when you don't pick a model yourself.
- Context-aware — the panel sends your currently open tabs (active file first) as context automatically.
- Review before write — every proposal is a diff you apply; the agent never edits a file directly.
- Rate and flag — every response can be rated, and a governance flag routes a concerning suggestion for review.
Limits & honest scope
- A BYOK model that doesn't have its key in the vault (or in env) won't respond — it shows as "needs a key" in the Models tab rather than failing silently mid-chat.
- Local models need the Creytix Model Server running (
npm run creytix:models:serve, or thecom.creytix.model-serverlaunchd service). It provisions its own engine and loads the weights already on disk — no separate install or model pull. - Not every listed model is production-trusted for the main edit — several are "parallel-only" or "candidate" trust, meant for cross-checks and fan-out drafts rather than the primary change.
- Provider coverage and per-model trust levels are still being reviewed; check the Models tab for current status rather than assuming this table is exhaustive.