# Darce

> Darce is an open source (MIT) AI coding agent for the terminal: the coding agent you can undo. It reads your codebase, edits files and runs commands using any of 250+ coding models via OpenRouter (Claude, GPT, Gemini, Qwen, DeepSeek, Kimi, GLM, Grok). Every change, including what shell commands did, can be reversed with /undo, and risky commands ask first with the reason.

Darce is built in public by one developer, Amer Sarhan. npm package: darce-cli. Requires Node.js 22+. Runs on macOS, Linux and Windows.

## Install

- Try with no account: `npx darce-cli` (first run offers a free trial, no sign-up)
- Install globally: `npm install -g darce-cli`, then run `darce` in a project folder
- Free account: `darce signup` (a daily usage allowance, no card; keeps trial history)
- Non-interactive / CI: `darce -p "task"` prints the result and exits
- Pick a model: `darce --model qwen/qwen3-coder`

## Key features

- /undo reverses every change, including files that shell commands created, changed or deleted (git-backed checkpoints in private refs; never touches your branch, index or stash). /rewind or Esc twice scrubs through the whole session.
- Risk-scored approvals: every command is scored (read-only, changes the project, reaches outside, destructive). In the default auto mode, risky commands ask first with the reason. Modes: auto, ask, plan (read-only), full.
- Second look at scripts: `npm run x`, `make y` and similar get checked by a fast classifier (TypeSafe Jev) for effects outside your machine; effects /undo can't reverse (deploys, migrations, releases) always ask.
- /swarm splits a task into 2-4 parallel agents, each in its own git worktree, then merges; /undo reverts the whole swarm.
- /derby races three models on the same task in separate worktrees; compare diffs, test results and cost, keep the best.
- 250+ coding models via OpenRouter; switch mid-task with Shift+Up/Down; /critic has a model from another vendor review every edit.
- WHY notes explain non-obvious changes; /resume continues past conversations; skills (SKILL.md, Claude Code skills work as-is); memory across sessions; image generation; paste screenshots.
- Secrets: credential-like environment variables are hidden from commands; known secret formats are redacted before reaching a model.

## Pricing

- Try it now: free trial, no account
- Free: $0, a daily usage allowance, models up to $5 per million output tokens
- Builder: $15/month, about 7x the Free allowance reset weekly, models up to $20 per million output tokens, including Claude Sonnet and Opus
- Power: $65/month, about 4x the Builder allowance reset weekly, every model
- Usage is measured by the real cost of model calls, so fast models (Claude Haiku, Qwen3 Coder) stretch an allowance much further than large ones.
- Every plan gets every feature. Limits are hard: no overage bills. Cancel any time.

## Docs

- [Getting started](https://darce.dev/docs/getting-started): Install Darce, the open source AI coding agent for your terminal, try it free with no account, and run your first task.
- [Undo](https://darce.dev/docs/undo): How Darce's /undo and /rewind reverse every change, including files that shell commands created, changed or deleted, using private git refs.
- [Safety](https://darce.dev/docs/safety): How Darce scores every command before it runs, which approval modes exist, how scripts get a second look, and how secrets are kept from models.
- [Swarm and derby](https://darce.dev/docs/swarm-and-derby): /swarm splits a task across parallel agents in git worktrees and merges the result; /derby races three models on one task so you keep the best diff.
- [Models](https://darce.dev/docs/models): Use 250+ coding models from Claude, GPT, Gemini, Qwen, DeepSeek, Kimi, GLM and Grok in Darce. Pick, search and switch models mid-task.
- [Commands](https://darce.dev/docs/commands): Every Darce slash command, command-line flag and keyboard shortcut: /undo, /rewind, /swarm, /derby, /model, /mode, /resume and more.
- [Configuration](https://darce.dev/docs/configuration): Every ~/.darcerc option, project instruction files, SKILL.md skills (Claude Code skills work as-is) and memory in Darce.

## Guides

- [When an AI agent deleted a production database: incidents and guardrails](https://darce.dev/ai-agent-deleted-production-database): A sourced timeline of AI agents deleting production databases (Replit, Claude Code with Terraform, Cursor at PocketOS) and the guardrails that prevent it.
- [Claude Code checkpoints vs other AI coding agent checkpoints](https://darce.dev/ai-coding-agent-checkpoints): How Claude Code checkpoints, Cursor restore checkpoint, Gemini CLI checkpointing, Codex, Cline, OpenCode and Aider undo work, and what each can't restore.
- [Best CLI coding agents in October 2026: 12 terminal agents compared](https://darce.dev/best-cli-coding-agents): Best CLI coding agent for you, October 2026: Claude Code, Codex CLI, OpenCode, Antigravity CLI, Pi, Darce and more compared on price, undo and models.
- [Claude Code with other models: router, OpenRouter, DeepSeek and local](https://darce.dev/claude-code-other-models): Run Claude Code with DeepSeek, GLM, OpenRouter, LiteLLM, claude-code-router or a local model: what Anthropic supports, and a simpler multi-model path.
- [Cursor alternatives: IDE and terminal options compared](https://darce.dev/cursor-alternatives): Best Cursor alternatives in 2026: Copilot, Zed, Cline, Kilo Code and Antigravity as editors; Claude Code, Codex CLI, OpenCode and Darce in the terminal.
- [OpenCode alternatives: 8 terminal coding agents compared](https://darce.dev/opencode-alternatives): The best OpenCode alternatives in 2026: Claude Code, Codex CLI, Kilo CLI, Crush, Aider, Gemini CLI and Darce, compared on models, pricing and local models.
- [Claude Code sandbox vs undo: what each protects in an AI coding agent](https://darce.dev/sandbox-vs-undo): How the Claude Code sandbox, Codex and Gemini CLI sandboxes and dev containers work, Windows and WSL support, what they miss, and why you need undo.
- [What is a coding agent? A plain-English guide](https://darce.dev/what-is-a-coding-agent): What is a coding agent? An AI that reads your code, edits files, runs tests and checks results in a loop. How it works, the kinds, the risks, how to pick.
- [Claude Code YOLO mode (--dangerously-skip-permissions) with a safety net](https://darce.dev/yolo-mode): What claude --dangerously-skip-permissions really does, Claude Code permission modes, Codex and Gemini YOLO flags, real incidents, and how to recover.
- [Full text of every doc, guide and comparison](https://darce.dev/llms-full.txt)

## What /undo can and cannot reverse

- Reverses: files Darce edited or created; files a shell command created, changed or deleted (rm, mv, formatters, generators); package.json and lockfile changes from an install; a whole /swarm merge as one step.
- Limits: shell-effect undo needs a git repository; gitignored files (like node_modules) are not snapshotted; undo history lasts for the session.
- Cannot reverse anything outside your machine: deploys, database migrations, git push, API calls. Darce asks before those.

## Comparisons

- [Claude Code alternative](https://darce.dev/alternatives/claude-code): Darce vs Claude Code, honest comparison and when to pick which
- [Claude Code rewind vs Darce undo](https://darce.dev/claude-code-rewind): what Claude Code's /rewind can't restore (shell command changes) and how to recover it
- [Darce vs OpenCode](https://darce.dev/compare/opencode): Darce vs OpenCode: two open source terminal AI coding agents compared on models, pricing, undo, approvals and parallel agents.
- [Darce vs Aider](https://darce.dev/compare/aider): Darce vs Aider: an autonomous coding agent and a git-native AI pair programmer compared on workflow, models, undo and pricing.
- [Darce vs Gemini CLI](https://darce.dev/compare/gemini-cli): Darce vs Gemini CLI: Google's Gemini-only terminal agent and a 250+ model open source agent compared on models, pricing and undo.
- [Darce vs Codex CLI](https://darce.dev/compare/codex-cli): Darce vs OpenAI Codex CLI: compared on models, pricing, sandboxing vs risk-scored approvals, and undo for shell commands.
- [Cline vs Claude Code vs Darce (2026)](https://darce.dev/compare/cline): Cline vs Claude Code, plus Darce: Plan/Act modes, bring-your-own keys vs subscriptions, checkpoints, command approvals and the Cline CLI, checked Oct 2026.
- [Cursor vs Claude Code vs Darce (2026)](https://darce.dev/compare/cursor): Cursor vs Claude Code, with Darce as a third pick: editor vs terminal, models, pricing, checkpoints, command safety and parallel agents, checked Oct 2026.
- [Kilo Code vs Claude Code vs Darce (2026)](https://darce.dev/compare/kilo-code): Kilo Code vs Claude Code, plus Darce: VS Code and CLI, 500+ models via Kilo Gateway vs Claude only, pricing, checkpoints and approvals, checked Oct 2026.
- [Pi vs Claude Code vs Darce (2026)](https://darce.dev/compare/pi): Pi coding agent vs Claude Code, plus Darce: what Pi is, whether it's safe (no built-in permission prompts), models, undo and pricing, checked Oct 2026.

## Models

- [Model directory](https://darce.dev/models): context windows, prices and the darce --model command for popular coding models
- [Claude Opus 5.5](https://darce.dev/models/claude-opus-5-5): `darce --model anthropic/claude-opus-5.5`
- [Claude Sonnet 5.5](https://darce.dev/models/claude-sonnet-5-5): `darce --model anthropic/claude-sonnet-5.5`
- [Claude Haiku 5.5](https://darce.dev/models/claude-haiku-5-5): `darce --model anthropic/claude-haiku-5.5`
- [Claude Fable 5.1](https://darce.dev/models/claude-fable-5-1): `darce --model anthropic/claude-fable-5.1`
- [Claude Opus 5](https://darce.dev/models/claude-opus-5): `darce --model anthropic/claude-opus-5`
- [Claude Sonnet 4.5](https://darce.dev/models/claude-sonnet-4-5): `darce --model anthropic/claude-sonnet-4.5`
- [GPT-6.1 Sol](https://darce.dev/models/gpt-6-1-sol): `darce --model openai/gpt-6.1-sol`
- [GPT-6 Sol](https://darce.dev/models/gpt-6-sol): `darce --model openai/gpt-6-sol`
- [GPT-6 Luna](https://darce.dev/models/gpt-6-luna): `darce --model openai/gpt-6-luna`
- [GPT-6 Astra](https://darce.dev/models/gpt-6-astra): `darce --model openai/gpt-6-astra`
- [GPT-5.6 Sol](https://darce.dev/models/gpt-5-6-sol): `darce --model openai/gpt-5.6-sol`
- [GPT-5.5](https://darce.dev/models/gpt-5-5): `darce --model openai/gpt-5.5`
- [GPT-5.3-Codex](https://darce.dev/models/gpt-5-3-codex): `darce --model openai/gpt-5.3-codex`
- [gpt-oss-120b](https://darce.dev/models/gpt-oss-120b): `darce --model openai/gpt-oss-120b`
- [Gemini 3.1 Pro Preview](https://darce.dev/models/gemini-3-1-pro-preview): `darce --model google/gemini-3.1-pro-preview`
- [Gemini 3.8 Flash](https://darce.dev/models/gemini-3-8-flash): `darce --model google/gemini-3.8-flash`
- [Gemini 3.5 Flash Lite](https://darce.dev/models/gemini-3-5-flash-lite): `darce --model google/gemini-3.5-flash-lite`
- [Gemma 4 31B](https://darce.dev/models/gemma-4-31b-it): `darce --model google/gemma-4-31b-it`
- [Qwen3 Coder 480B A35B](https://darce.dev/models/qwen3-coder): `darce --model qwen/qwen3-coder`
- [Qwen3 Coder Next](https://darce.dev/models/qwen3-coder-next): `darce --model qwen/qwen3-coder-next`
- [Qwen3 Coder Plus](https://darce.dev/models/qwen3-coder-plus): `darce --model qwen/qwen3-coder-plus`
- [Qwen3 Coder Flash](https://darce.dev/models/qwen3-coder-flash): `darce --model qwen/qwen3-coder-flash`
- [Qwen3 Coder 30B A3B Instruct](https://darce.dev/models/qwen3-coder-30b-a3b-instruct): `darce --model qwen/qwen3-coder-30b-a3b-instruct`
- [Qwen3.8 Max (0902)](https://darce.dev/models/qwen3-8-max-0902): `darce --model qwen/qwen3.8-max-0902`
- [Qwen3.8 Flash](https://darce.dev/models/qwen3-8-flash): `darce --model qwen/qwen3.8-flash`
- [DeepSeek V4 Pro 0423](https://darce.dev/models/deepseek-v4-pro): `darce --model deepseek/deepseek-v4-pro`
- [DeepSeek V4.1 Flash](https://darce.dev/models/deepseek-v4-1-flash): `darce --model deepseek/deepseek-v4.1-flash`
- [DeepSeek V3.2](https://darce.dev/models/deepseek-v3-2): `darce --model deepseek/deepseek-v3.2`
- [Kimi K3](https://darce.dev/models/kimi-k3): `darce --model moonshotai/kimi-k3`
- [Kimi K2.7 Code](https://darce.dev/models/kimi-k2-7-code): `darce --model moonshotai/kimi-k2.7-code`
- [Kimi K2.6](https://darce.dev/models/kimi-k2-6): `darce --model moonshotai/kimi-k2.6`
- [Kimi K2 Thinking](https://darce.dev/models/kimi-k2-thinking): `darce --model moonshotai/kimi-k2-thinking`
- [GLM 5.3](https://darce.dev/models/glm-5-3): `darce --model z-ai/glm-5.3`
- [GLM 5.3 Flash](https://darce.dev/models/glm-5-3-flash): `darce --model z-ai/glm-5.3-flash`
- [GLM 5.1](https://darce.dev/models/glm-5-1): `darce --model z-ai/glm-5.1`
- [Grok 4.7](https://darce.dev/models/grok-4-7): `darce --model x-ai/grok-4.7`
- [Grok Build 0.1](https://darce.dev/models/grok-build-0-1): `darce --model x-ai/grok-build-0.1`
- [MiniMax M3](https://darce.dev/models/minimax-m3): `darce --model minimax/minimax-m3`
- [MiniMax M2.7](https://darce.dev/models/minimax-m2-7): `darce --model minimax/minimax-m2.7`
- [Devstral 2 2512](https://darce.dev/models/devstral-2512): `darce --model mistralai/devstral-2512`
- [Codestral 2508](https://darce.dev/models/codestral-2508): `darce --model mistralai/codestral-2508`
- [Mistral Large 4](https://darce.dev/models/mistral-large-4-0): `darce --model mistralai/mistral-large-4-0`

## Links

- [Website](https://darce.dev)
- [Changelog](https://darce.dev/changelog)
- [GitHub](https://github.com/AmerSarhan/darce-cli)
- [npm](https://www.npmjs.com/package/darce-cli)
- [Discord](https://discord.gg/u447rt6Xfq)