Yes, the CLI is MIT licensed and on GitHub. It does not send telemetry.

Darce — the AI coding agent you can undo

Shipped recently

0.19.0Oct 31
Upgrading takes one key
0.18.0Oct 20
Your conversation is saved after every change, not only when a turn ends
0.17.2Oct 18
`/clear` no longer erases the conversation you cleared
0.17.1Oct 15
Fewer pointless approval prompts

Undo covers what the shell did, too.

Darce snapshots your project before every step, not just before its own edits. A folder rm deleted, a file mv moved, a lockfile an install rewrote: one /undo and it's back. Your branch, index and stash are never touched.

This session, step by step/rewind · esc esc

Edit
src/db.ts
Write
src/cache.ts
Bash · rewind here
rm -rf test
Bash
npm install zod
Edit
src/app.ts

Rolls back

Files Darce edited or created
Files a shell command deleted, moved or changed
package.json and lockfile changes from an install
Any earlier step, not just the last one

Asks first, because it can’t

Database migrations and writes
Deploys and git push
API calls that change something
Anything outside this project

It only asks when it can’t undo.

Every command is scored before it runs. If /undo can reverse it, it just runs. If it reaches outside your project or your machine, Darce stops and tells you why. Scripts with harmless names get read first.

CommandWhat it would doDarce

git statusOnly readsRuns

npm testChanges nothing /undo can’t reverseRuns

npm run migrateWrites to a databaseAsks first

node scripts/sync.jsHarmless name; a second check reads the script and finds it POSTs customers to a CRMAsks first

rm -rf ~/projectsDeletes files outside this projectAlways asks

Allow Bash?asks first

node scripts/sync.js

POSTs data to an outside API; /undo can't reverse it

y allow oncen denyesc deny and stop

And a few things that make it different.

See what every change touches

/brain opens a live view in your browser: the file Darce is editing, what it imports, what imports it, and the exact diff.

Race three models, keep the best

/derby runs one task on three models in separate worktrees. Read each answer as it lands and apply the one you'd merge.

▸1claude-sonnet-5.5+18 −613s

2gpt-5.6-sol+22 −919s

3qwen3-coder+31 −129s

enter apply · esc discard all

Each model works in its own git worktree, so nothing collides. Your files stay untouched until you pick one, and /undo still works after.

Split big work across agents

/swarm plans the parts, runs each in its own worktree at once, and merges them back in one step.

> /swarm add stats, CSV export and validation

Completion stats moduledone

CSV exportdone

Title validationdone

Waiting for all three…

Hear what it's doing

/voice on: a short spoken update when a task runs long, when it needs you, and when it's done.

“Before I push this, can I run the migration? It can't be undone.”

It explains the bug

When a fix depends on something the code doesn't show, Darce adds one short note so you spot it next time.

WHYforEach ignores the promises its callback returns, so nothing waited.

Any top model. One account, no API keys.

Switch mid-task with Shift ↑↓. You stop at your limit and are never charged extra.

Qwen3 CoderFree
Claude Sonnet 5.5Builder
Claude Opus 5.5Builder
DeepSeek V4.1 FlashFree
GPT-5.6 SolBuilder
Gemini 3.1 ProBuilder
Gemini 3.8 FlashFree
Kimi K3Builder
Claude Haiku 5.5Free

And 240 more, filtered to models that can reliably use tools. Browse them all

Free

$0

Try it with no account, then a daily allowance. Fast models like Claude Haiku and Qwen3 Coder.

BuilderRecommended

$15/ month

About 7× the Free allowance, weekly. Adds Claude Sonnet and Opus, GPT and Gemini Pro.

Power

$65/ month

About 4× Builder, weekly. Every model, including the priciest reasoning ones.

Usage is measured by what the work costs, so fast models go further. Every plan has every feature: undo, approvals, derby, swarm, brain. Full pricing

Darce is coming to the Mac.

The same agent and the same undo, in an app with your files and an editor beside it. Choose how much Darce explains as it works, from none at all to a lesson.

acme-todos

▶src
▶components
TSTodoList.tsx
TSTodoItem.tsx
▶hooks
TSuseTodos.tsM
▶lib
TSapi.ts
TStypes.ts
▶tests
TSuseTodos.test.ts
{}package.json
¶README.md

Checkpoints⌘Z

Step 4 · just nowUndo

acme-todos/main

ShipUnderstandLearn

Claude Haiku 5.5

TSuseTodos.tsTSapi.ts

src › hooks › useTodos.ts

1import { useCallback, useRef, useState } from 'react'

2import { saveTodo } from '../lib/api'

3import type { Todo } from '../lib/types'

4

5export function useTodos(initial: Todo[]) {

6 const [todos, setTodos] = useState(initial)

7+ const queue = useRef<Promise<void>>(Promise.resolve())

8

9 const save = useCallback((todo: Todo) => {

10 setTodos(ts => ts.map(t => t.id === todo.id ? todo : t))

− void saveTodo(todo)

11+ // One save at a time, so a slow request

12+ // can't land after a newer one.

13+ queue.current = queue.current.then(() => saveTodo(todo))

14+ return queue.current

15 }, [])

16

17 return { todos, save }

18}

DarceDone in 14s

Editing a todo twice quickly sometimes keeps the old text. Can you fix it?

read hooks/useTodos.ts

read lib/api.ts

edit hooks/useTodos.ts+5 −1

$ npm test18 passed

Fixed. Two saves were racing each other.

WHYsaveTodo wasn't awaited, so two quick edits sent two requests at once. When the first was slower, it landed last and the server kept the old text.

I didn't debounce the input instead: that delays every save and still races on a slow network.

Ask Darce…

Understand · Haiku 5.5⌘↵

Ln 13, Col 5 · TypeScriptStep 4 of 4 · undo ready

Builds, then explains each decision and what it ruled out.

Why it's called Darce

Darce comes from درس, dars: Arabic for a lesson.

I started it because I wanted an agent I could let run without watching every step, and that only works if every step can be taken back. That's why undo is the core.

But the name says where it's going. An agent that writes your code can quietly leave you understanding less of it. I want the opposite: a tool that teaches you while it works, so you come out of every session knowing a little more. The WHY notes are the first piece of that. There's more coming.

I build Darce on my own, in public, and it's in early beta. If something breaks, or you have an idea for how it should teach, tell me on GitHub or Discord. I read everything.

Amer Sarhan

Questions

How is usage measured?

Each plan has an allowance: Free per day, Builder and Power per week. It's measured by what the work actually costs, so fast models like Claude Haiku and Qwen3 Coder stretch it much further than Opus. Darce shows how much you've used, never a bill. Tasks per plan, by model.

What happens when I run out?

Darce stops and tells you when your allowance resets. You're never charged extra: plans have a hard limit. A cheaper model makes what's left go further.

What can’t /undo reverse?

Anything that leaves your computer: database writes, deploys, git push, payments, API calls that change something. That's exactly why those always ask first. Everything inside your project, including what shell commands did, rolls back. How undo works.

Where does my code go?

Darce runs on your computer. To answer, it sends your request and the parts of files it reads to api.darce.dev, which passes them to the model provider you picked through OpenRouter. Darce doesn't store your code or prompts; it keeps your email, plan and how much you've used. Privacy policy.

Do I need my own API key?

No. One Darce account reaches 250+ models. You never paste a provider key.

Which models does each plan include?

Free covers fast, low-cost models: Qwen3 Coder, DeepSeek, Gemini Flash, Claude Haiku. Builder adds the big ones, including Claude Sonnet and Opus, GPT and Gemini Pro. Power adds the most expensive reasoning models and about four times Builder's allowance. Every model by plan.

How is it different from Claude Code?

Claude Code is excellent, and Darce borrows its conventions: / commands, @ files, Esc to stop, Shift+Tab modes, CLAUDE.md. The differences: Darce's undo also reverses what shell commands did (Claude Code's checkpoints don't track Bash), and it works with any model, not just Claude. Full comparison.

Is it open source?

Yes, the CLI is MIT licensed and on GitHub. It doesn't send any telemetry.

The coding agent you can undo.