Godspeed runs Claude Code, Codex, Cursor and the rest in one native Mac window. It opens in 133 ms, idles at 12 MiB, and works with the subscriptions you already pay for.
Apple silicon. Free and MIT-licensed. No account needed. All downloads
Runs what you already use
Godspeed drives the agent CLIs you already signed in to, each over its own protocol. It never resells model access, so your plans and limits stay yours.
Built in
One click from the ACP registry
Benchmarks
Godspeed's headless engine against the T3 Code 0.0.40 server. Medians of six interleaved trials on an M2 MacBook Air.
These compare the headless servers, not the apps, and measure no agent or model work. The machine wasn't idle during the run, and the two search calls do different work. Against Zeron, which Godspeed is forked from, cold start is the same 65 ms and Zeron idles about 1.6 MiB lighter. Method and raw data: vs T3 Code, vs Zeron.
Only here
Five things we checked against both codebases before claiming them.
Every agent on every device, on one board.
Cmd+Shift+K opens a board with each agent's status, elapsed time, latest activity, branch and diff counts. Move with the arrow keys, open a chat with Return, approve the low-risk ones with A.
The safe ones go with one key.
Every approval and question from every agent lands in one Needs-you inbox. Godspeed reads the actual command: read-only commands and edits inside the chat's own folder are low risk. Network calls, git push, .env files and anything outside the repo are high, and always show you the full command.
Every turn's tokens, on a receipt.
Each finished turn prints a receipt: wall time, tool calls by kind, files changed, and tokens in, out and cached. On API-billed accounts it adds an estimated cost. Copy it as text or as an image.
Schedule, steer and approve from your phone.
Zoomies are agents that run on a schedule, a webhook or a file change. The engine serves a Telegram bot, so you can schedule one, run it now, pause it and answer its approvals from Telegram, even when the engine lives on a server.
/schedule weekdays 18:00 Summarise new PRs
Scheduled #3: weekdays 18:00 (America/Los_Angeles).
Next run Wed 28 Oct 18:00
Manage it with /schedules.
Approval needed · Add search to notes
Run command:
git push origin feat/search
Allow once
Your Dock icon runs with your agents.
The speedster jogs while agents work, sprints at four or more, waves when something needs you, celebrates a finish and naps after ten quiet minutes. Try it here. Settings → Appearance turns it off.
Jog 1 to 3 agents working
Checked against T3 Code (origin/main 40d335a7) and Zeron (upstream/main 8b7afb30) source on 28 October 2026. Scheduled agents and agents driving the preview browser exist in T3 Code too, so they aren't on this list.
Also shipping
One long-lived chat plans the work and steers sub-agents. Each one gets a card, and the parent wakes when a child finishes, fails or needs you.
Hit the bolt in the new-chat composer, type "weekdays at 6pm", see the next run, send.
Install the headless engine on a Linux box over your own SSH or Tailscale. Agents keep going with the laptop shut.
Copilot CLI, Factory Droid, Gemini CLI and Kimi install in one click, each pinned to a checksum.
Named instances with their own config dir, env and account. Secrets go to the keychain.
Clone a project onto your server or another Mac and bring the .env files git doesn't track, through your own signed-in engines.
Sub-agents can start in a fresh worktree on their own branch. Cleanup never touches unpushed work.
Approve, comment, resolve threads and merge from the PR pane, with GitHub's reason when it refuses.
Failed checks, new review comments and merge conflicts wake the PR's chat. Nothing is approved for you.
Live 5-hour and weekly limits for every account, a tachometer in the composer, and history at API prices. Limited chats resume when the limit resets.
Downloads, checks the signature and notarization, then asks you to restart. Never a downgrade.
Parallel agents share one jobserver. In testing, peak memory went from 4.7 GB to 2.0 GB.
Follow your Mac's chats, approve, and dictate from your phone. Build it yourself for now.
Import chats from T3 Code, Claude Code, Codex and Cursor. Claude Code and Codex sessions carry on.
No account needed. Chats live on your Mac, searchable from a local index.
From the launch thread
133ms is wild. I built my own Neovim-like editor in Rust (Nevi) and that startup and memory gap is a big reason I stopped wanting Electron for daily tools.
Anthony Amaro@Annthonny
The receipt after each turn is my favorite part, love seeing what a run actually cost.
Denys@dkolosovs
reading the actual command instead of vibes is the right call. push, .env and anything outside the repo always being high is exactly where i'd draw the line too
ZeroX@dxiaolong
12mb idle with that many agents is wild.
Rohan Bhanot@RohanBhanotAI
You're doing god's work
Sophie@Sovex666
Replies to the launch post on X, quoted as written (emoji left out). Each links to the original.
shrithan, @DevaiahShrithan
Star on GitHub 12Every dot is a release, pulled live from GitHub.
Plainly
No pricing page, because there's no price.
Download
v0.5.9, released 28 Oct 2026. The Mac app is signed and notarized by Apple, and updates itself from there.
New, lightly tested. Also runs headless.
Not on TestFlight yet. Build it with Xcode.
Not yet.
