AI assistant
Describe what you need in natural language — AI produces a typed intent, not HCL directly, by default.
Chat replaces a blank Terraform file. Describe what you need in plain language and the assistant extracts a typed, structured intent — resource type, provider, key attributes — for you to review before anything is generated. By default, the language model doesn't write HCL directly.
- Doesn't generate raw HCL by default — a typed, structured request comes first
- Understands intent across all 12 supported providers
- You review and approve the extracted intent before code is generated
How it works
Every chat turn runs a fixed pipeline: guardrails and budget checks, an intent router that extracts a typed request, context assembly (schema retrieval, RAG, tenant memory), model selection with automatic failover, then a tool loop that drafts the request under a per-thread lock. For Terraform providers, a Planner and Terraform agent prefetch a plan before the main loop runs — and fail closed, asking you a question, if the model can't produce one confidently.
When the model can't produce a confident answer — a timeout, an invalid tool call — the orchestrator doesn't fall back to a silent heuristic. It asks a clarifying question instead, and the legacy heuristic fallback only re-activates if you explicitly opt into it.
A real SSE event pair from a chat turn
event: tool_use
data: {"type":"draft_resource","input":{"resourceType":"aws_instance","provider":"AWS"}}
event: proposed_defaults
data: {"category":"instance_type","value":"t3.medium","source":"catalog_default"}