You read the code
Every change an agent makes shows up as a numbered diff. Keep or undo it by hunk, by file or all at once, and nothing counts as done until you have.
Agents write more of the code now. You still read it, review it, and some days you just want to write it yourself. Ember is a real editor for running many Claude agents at once: every task on its own branch and pull request, every change yours to keep or undo. It runs on the Claude subscription you already pay for.
A lot of agent tools ask you to stop looking at the code: type what you want into a box and trust what comes back. Ember is built on the opposite idea. You are the engineer. The agents work for you, in an editor where you can see, change and overrule everything they do.
Every change an agent makes shows up as a numbered diff. Keep or undo it by hunk, by file or all at once, and nothing counts as done until you have.
A real editor on VS Code's open-source base, with your extensions, next-edit predictions as you type and Cmd+K for the lines you want changed your way.
Branches, commits, pull requests and merges are plain git that you can see and redo in a terminal. Merging stays with you.
Plan before anything is built, steer a chat while it works, rewind to any message with your files put back, or take a branch over and hand it back.
One board of every branch in flight, with its state and one next step. No reading five transcripts to find out what five agents did.
Your CLAUDE.md, skills, hooks, subagents and MCP servers are the same plain files the claude CLI reads. Nothing is locked inside Ember.
One agent in one terminal is easy to follow. Five, each with a pull request and CI, means knowing which branch is where, what each is waiting for and what to do next, without reading five transcripts. Ember follows eight rules on every screen, so you learn them once.
Claude never commits on your base branch. It makes a branch before its first commit, and parallel work goes in a git worktree of its own.
Agents work in background worktrees while your own tab stays exactly as you left it. When a branch is done, you choose what happens to it.
No shadow copy of your work. Bringing changes over is git merge --squash and undoing it is git reset --merge. The claude CLI sees the same state.
A worktree is removed only when everything in it is on its branch. A branch is deleted only when git shows its work landed.
Every branch is Needs you, Working, Waiting, Ready, Merged or Idle, and it offers one next step. Every screen reads the same list.
Needs you is ranked by what each item frees up. A question holding up three stacked branches comes before a PR that only waits.
Where PRs go, which branches are protected, how branches are named, the pre-PR check. Ember detects what it can and shows where each value came from.
Agents push their own branch and open their own PR. Anything into a protected branch needs your words, and hooks are never skipped.
Ember opens on what needs you first, ranked by what it unblocks: a command to approve, a question to answer, a pull request to merge. Then what's working, your plan limits and the last 24 hours. Every number opens the board on just those branches.
The composer's Where pill sends a chat to this tab, a new branch, a background worktree from a base it fetches first, or a branch that already has a worktree or a PR. Claude never commits on your base branch.
Your messages on the right, a short diff on each edit, and every branch, commit, push, pull request and CI result as one quiet line. Queue or steer while Claude works, rewind to any message with your files put back, or fork the chat to try another way.
A waiting command offers Run, Skip or Always run, with the exact rule and where it's saved. Commands that can't be undone always ask, and Ember's safety checks fail closed.
Each to-do says what it changes and what it waits for, so Bonfire knows which can run together. Every to-do gets its own worktree, chat, branch and pull request, a reviewer checks it, and passing work merges in order.
A numbered diff with Keep or Undo per hunk, per file or all of it, and Agent Review findings anchored under the lines they're about. Fix with Claude sends a finding back to the chat that wrote the code.
Send one prompt to two to four models at once, each on its own branch from the same base. Compare shows their diffs side by side. Keep one, and Ember removes the others through the same checks as everything else.
The phone below is the real Ember phone app, running in this page.
An installable app that talks to Ember on your own computer. It opens on Now: what needs you, ranked, with the same one next step the Dashboard shows.
Each edit is one line, a question's choices and an approval's answers are full-width buttons, and send stays where your thumb expects it. Dictate a reply, or start a new chat.
Open the same address on an iPad. The window's width picks the layout: lists beside what you open, diffs beside the chat, and shortcuts for everything with a keyboard attached.
Tailscale, your local network, ngrok, a Cloudflare tunnel or your own URL. A 6-digit code that lasts five minutes pairs a device, each device gets its own token, and every tap is checked against the branch as it is right now.
Ember drives the local claude CLI through the Claude Agent SDK, so every chat runs on your Claude Pro or Max plan. Ember never asks for, reads or stores your Claude credentials, and it removes stray API keys from the agent's environment.
Next-edit predictions as ghost text with inline diffs and tab-to-jump, on a model you pick: a next-edit model locally or hosted, or any fill-in-the-middle model.
Cmd+K streams an edit into the file and stays open for a follow-up. In the terminal it writes the command without running it.
Hybrid keyword and semantic search, a codebase_search tool for Claude, every repository and worktree kept current, embeddings from the endpoint you choose.
GitHub and Jira issues in one table, each with its branch, its chat and its PR beside it. Start one and Ember names the branch from your template.
Git for this worktree: stage by file, hunk or line, a commit message Claude drafts, history with the branch graph, and a guide through a paused rebase.
Send one prompt to two to four models at once, each on its own branch from the same base. See the diffs side by side and keep the best.
Claude can navigate, click, type and read the console, and check UI changes visually against a baseline. You can take over at any time.
Ember tells you out loud when a chat needs you, by name, and you answer into that chat by voice. On your own speech endpoints or your OS's voice.
Work beside Personal: choose which account each folder, window or chat uses, and decide what happens when one runs low.
Live 5-hour and weekly limits in the composer, the Dashboard and the status bar, and a usage page by day, model, project and account.
Other subscriptions, any Anthropic-compatible gateway or API key, or local models on oMLX, Ollama or LM Studio. Keys live in the OS keychain.
Ember has no cloud of its own and no account to create. Your prompts and code go to Anthropic, or the model provider you choose, exactly as they do with Claude Code. Chat history, the index and usage stats are stored on your machine.
Your habits carry over: agent chat with Plan and Ask, parallel agents, Tab, Cmd+K, Keep and Undo, rules and a codebase index. What's different is underneath.
| Ember | Cursor | Claude Code extension for VS Code | |
|---|---|---|---|
| The agentWhat actually does the work | ✓ Claude Code itself, driven through the Claude Agent SDK | ◐ Cursor's own agent, with Claude among its models | ✓ Claude Code |
| What you pay for | ✓ The Claude subscription you already have. No API key. | — A separate Cursor subscription | ✓ Your Claude subscription |
| Your Claude Code setupCLAUDE.md, skills, hooks, subagents, plugins, MCP, permissions | ✓ All of it applies, each with a real settings UI. One configuration for the terminal and the editor. | — Its own rules and settings | ✓ Applies; mostly edited as files |
| Many agents at onceKnowing which branch is where and what it waits for | ✓ A board of every branch in every window: six states, one next step each, ranked by what it unblocks | ◐ Parallel agents, listed by agent | ◐ Several sessions, one panel each |
| Multi-agent plan buildsOne plan, several agents, dependencies respected | ✓ Bonfire: a branch and PR per to-do, stacked PRs, merge order, a usage forecast, scheduled runs | — | — |
| Review before it lands | ✓ Keep or Undo per hunk, file or everything; Agent Review findings under the line; Fix with Claude; Agent Blame | ✓ Per-change accept and reject, with review tools | ◐ Accept or reject each proposed edit |
Based on our own use of each product and their public documentation, as of November 2026. Products change; if something here is out of date, tell us and we'll fix it.
Our rules, our shortcuts and our daily habits were all shaped by Cursor. When it stopped being a viable answer for us, we moved our work to Claude Code.
Claude Code is an excellent agent. But the VS Code extension left a lot to be desired for the way we work: no next-edit predictions, no inline edit, no hunk-by-hunk review, and no way to see five agents on five branches at once.
So we built the editor we wanted around the agent we trust, on the subscription we already had. Then we kept going, onto our phones.
Read the whole storyDevelopers who know how to code and still want to. Ember is a full IDE, not a chat box that replaces one: you get a real editor with next-edit predictions and inline edit for the code you write yourself, a review tab where you keep or undo every change an agent makes, and git that stays yours. The agents do more of the typing; you stay the engineer.
Ember is a desktop IDE built around Claude Code. It is a complete editor (built on VS Code's open-source base) with Claude's agent, a board of every branch your agents are working on, multi-agent plan builds, a review tab, next-edit predictions and a phone app. It runs on macOS, Windows and Linux.
No. Ember runs on your Claude Pro or Max subscription through the Claude Code CLI on your computer. It never asks for, reads or stores your Claude credentials, and it removes stray ANTHROPIC_API_KEY and ANTHROPIC_AUTH_TOKEN variables from the agent's environment so a chat is never billed per token by accident.
The agent in Ember is Claude Code itself, on your own Claude subscription, and your rules, skills, hooks and MCP servers are the plain files Claude Code already reads. Parallel work is organized around git: one task, one branch, one pull request, with a board that shows each branch's state and one next step. The habits carry over: agent chat with Plan and Ask, parallel agents, Tab, Cmd+K, and Keep or Undo per change.
The extension adds a Claude panel to an editor. Ember is a whole editor designed for running many Claude agents: a branches board and Dashboard, Bonfire plan builds with several agents at once, a review tab with Keep and Undo per hunk and Agent Review findings, next-edit Tab predictions, Cmd+K inline edit, a codebase index, live usage limits and a phone app. It uses the same Claude Code configuration, so the terminal, the extension and Ember all behave the same way.
Yes. Ember has an installable phone app that connects to Ember on your own computer. It opens on Now: what needs you, ranked, each with one button. You can answer a question, approve a command, merge a pull request, remove a worktree, stop an agent or start a new chat. You choose how the phone reaches your computer: Tailscale, your local network, ngrok, a Cloudflare tunnel or your own URL.
You need the Claude Code CLI, signed in with your Claude account. Already signed in? Open Ember and you're done.
Unzip and move Ember.app to Applications. It's notarized, so it opens without a warning, and it updates itself in place.
The installer is for your user only (no admin) and can add ember to your PATH. It isn't code-signed yet, so SmartScreen asks first: More info, then Run anyway.
The .deb and .rpm add Ember to your app menu, put ember on your PATH and handle ember:// links. Or extract the tar.gz and run ember.
Download it and use it. If Ember earns a place in your day, you can support its continued development for $5, from inside the app.