Your terminal,
now autonomous.
Crayon is an AI coding agent that lives in your terminal. It reads your whole codebase, plans, runs commands, and self-heals its own errors — until the task is actually done.
Plug in any model · run it anywhere
The loop
It doesn't just autocomplete. It finishes the job.
Understand
Crayon indexes your repo with tree-sitter and a local SQLite graph, then pulls only the handful of files a task actually needs — not your whole codebase into the prompt.
Plan
It drafts an explicit, step-by-step plan before touching anything. In plan mode you review and approve it — nothing runs until you say so.
Execute
It runs commands, writes surgical diffs, and edits files under the permission mode you set — Ask, Auto-Edit, or full Auto. You stay in control of the blast radius.
Self-heal
When a test breaks or a build fails, Crayon reads the error, patches its own work, and re-runs — looping until the task is verified done, not just attempted.
Capabilities
Everything an agent needs to ship real work.
Context Engine
A local tree-sitter + SQLite index resolves the exact files a task needs. Semantic search and dependency graphs — not your whole repo stuffed into the prompt.
Five model providers
Hot-swap between Anthropic, OpenAI, Google, OpenRouter, and local Ollama mid-chat — without leaving your flow.
Runs fully offline
Point Crayon at a local Ollama server for zero-cost, private runs. No API key, no data leaving your machine.
Permission modes
Ask, Auto-Edit, or Auto. Fine-grained control with guards against path traversal, oversized reads, and destructive commands.
Self-healing loop
Intercepts failing tests and broken builds, patches its own work, and re-runs until the task is verified done.
MCP extensible
Connect any Model Context Protocol server to give the agent new tools and integrations.
VS Code, too
A transparency-first panel with live reasoning, tool cards, and click-to-source path:line citations.
Auto-compaction
Intelligent history compaction keeps token burn low, with real-time cost tracking per session.
/color
Your terminal, your colors.
Crayon ships a whole box. Pick one — the whole page recolors, just like the /color command in the CLI.
Why Crayon
An agent you can actually trust with your codebase.
Every plan, command, and diff is shown before it runs. No black boxes — you see the reasoning and approve the blast radius.
Answers come from your actual code, resolved by the index — not guessed from a model that never read your repo.
Open source and MIT licensed. Run it on your models, your machine, your terms — including fully offline.
Pricing
Free, and open source.
Crayon itself costs nothing. You only ever pay your own model provider for tokens — or nothing at all when you run it locally with Ollama.
Everything, for everyone. MIT licensed.
- The full autonomous agent — no feature gates
- Context engine, self-healing, permission modes
- CLI + VS Code extension
- All five model providers & MCP servers
- Runs fully offline with Ollama
Model usage
Bring your own key for Anthropic, OpenAI, Google, or OpenRouter — you're billed by them at cost, with per-session tracking in the CLI.
local · private · no key required
Questions
Good to know
Autocomplete suggests the next line while you type. Crayon takes a whole task — reads the relevant code, writes a plan, runs commands, edits files, and fixes failing tests on its own, showing you each step along the way.
Instead of stuffing your entire repository into the prompt, Crayon indexes the codebase with tree-sitter and a local SQLite graph, then loads only the files, symbols, and imports a task needs. Lower cost, fewer hallucinations.
You choose the permission mode — Ask, Auto-Edit, or Auto. Edits are applied as reviewable diffs (with per-hunk approval), and there are built-in guards against path traversal, oversized reads, destructive commands, and SSRF on outbound fetches.
Anthropic, OpenAI, Google, and OpenRouter with your own key, or a local Ollama server for fully offline, zero-cost runs. You can hot-swap models mid-conversation.
Yes — Crayon is open source under the MIT license. You only pay your own model provider for tokens, or nothing when running locally with Ollama.