Skip to content

Corvid documentation

Corvid is a programming language where dangerous AI actions either compile with a proof or do not compile at all. The compiler reads approve, Grounded<T>, effect rows, and budget annotations the way other compilers read types, and it refuses to emit a binary if a tool call that should be gated by an approval is reachable without one.

The docs follow the Diataxis shape: tutorials, how-to guides, reference, explanation. Use whichever section matches what you are doing.

The Corvid Book — sequential learning

Read top to bottom to learn the language.

  1. Why Corvid
  2. Install
  3. Quickstart
  4. Tutorial: build a refund agent
  5. Syntax basics
  6. Types
  7. Modules and imports
  8. Effects
  9. Approve
  10. Grounded
  11. Agents
  12. Prompts
  13. Errors and Result
  14. Pattern matching
  15. Project layout
  16. Testing
  17. Building and targets
  18. Replay

Guides — task-focused how-to

Recipes — small focused patterns

RAG, approval-gated tool, multi-step agent with checkpoints, provider routing, local model fallback, audit log per decision.

Reference — lookup material

Migration — coming from another language

Operations — running Corvid in production

Security — model and policy

Internals — spec and implementation

Help

Meta — project-level docs

Phases — development records

Phase-by-phase dev records, audits, and refactor logs. These are historical engineering documents, not user-facing material. Browse phases/phase-NN-*.md for any phase mentioned in ROADMAP.md.