Caret is a free, open-source (MIT) AI code editor built for speed and privacy. The core is written in Rust on Tauri 2, and the interface runs in the system WebView with React and Monaco,
so there is no bundled Chromium and no Node runtime: the editor idles at around 110 MB of memory and paints its first frame in about 330 ms.
Everything you expect from a modern AI editor is built in. Tab gives multi-line ghost-text completions from any configured model, with partial accept and next-edit prediction. Cmd+K edits
or generates code inline with a streaming diff you accept or reject, and in the terminal it turns plain English into a shell command. The agent pane has Agent, Ask, Plan and Debug modes
with streaming tool calls, checkpoints, per-file review and a live cost readout, and every agent turn ends with a verified diff: Caret runs the project's checks and shows what passed
before you accept.
Bring your own keys or your own subscription. Caret talks directly to Anthropic, OpenAI, Google, OpenRouter, Ollama, LM Studio, vLLM and any OpenAI-compatible endpoint; nothing goes
through a Caret server. It also runs external agents over the Agent Client Protocol with the login you already have: Claude Code, Codex (ChatGPT plan), Cursor, Grok Build, Gemini CLI,
OpenCode and Pi, each with its model, effort and permission mode in the composer, and a second account per provider when you need one. Auto routing can send each request to the cheapest
of your own models that can handle it.
Safety is a first-class feature. Every shell command an agent wants to run, including commands from external agents, is parsed and classified before it executes. Blocked commands never
run, destructive ones always ask, and the rest run inside the macOS Seatbelt or Linux bubblewrap sandbox. Permission modes (Supervised, Auto-accept edits, Auto, Full access) can be set
per thread, per project and capped for a whole team.
Threads work the way long agent sessions need: pin, snooze, settle and archive; queue a message or steer the running turn; fork from any point; hand a thread to another agent or model; or
fan one prompt out to several models side by side, each in its own git worktree. A Usage page tracks tokens and cost across Caret and your external agents, Caret can act as an MCP server
so other agents can orchestrate it, and "caret serve" puts the full editor in a browser on your LAN or over Tailscale.
The rest of the editor is complete: an LSP client for TypeScript, Rust, Python, Go, C/C++, Java and more, a debugger speaking the Debug Adapter Protocol, a semantic terminal with command
blocks and clickable paths, source control with AI commit messages and a review mode, a local codebase index with optional embeddings, MCP servers, rules files (.caret, .cursorrules,
AGENTS.md, CLAUDE.md), hooks, memories, skills and subagents. Existing Cursor projects work unchanged.
No comments or reviews, maybe you want to be first?