generation 3 · agentic · macOS · MIT

The Atom is here

The hackable editor, reborn as an agentic, AI-native editor. Six agents in quiet orbit — they plan, edit, run, and verify. On your key, on your machine, no middleman.

no account · no telemetry · bring your own key

read_file edit_file run_command update_plan ask_user use_skill verify_loop checkpoint background_exec
L01agent runtime

Autonomous where it earns it.
Gated where it matters.

Every edit lands as a reviewable change. Every command waits for your approval. Before the agent says done, it diffs its own diagnostics and runs your tests — repairing what it broke, bounded and cancellable.

atom++ ai
you
Ship a dark-mode toggle across the app
atom++ ai

I'll scan the theme layer first to see which tokens exist.

list_files src/**/*.{ts,tsx}
read src/theme.ts
atom++ ai

Light tokens only. Adding a dark palette and wiring the toggle:

✎ edit src/theme.ts +18 2Keep ✓ · Undo
▶ run npm testapproved by you
✗ 1 failing — contrast below AA · patching…
✎ edit src/theme.ts +6 1Keep ✓ · Undo
▶ run npm test
✓ 31 passed · 1.1s
⛨ verify — 0 new diagnostics · objective complete
↩ checkpoint saved — restore this turn anytime
👍👎Haiku 4.5
+ Add context
Plan scan theme.ts → add tokens → wire toggle → verify
</> AgentHaiku 4.5 ▾
apply-then-review

Keep / Undo on every file

Edits land instantly and await your verdict — per file or all at once, live +/− totals, state survives reloads.

⛨ self-verification

Proves itself before reporting

Diffs diagnostics against a pre-run baseline and runs your verify command; new failures loop back for repair — bounded, never runaway.

↩ checkpoints

Rewind any turn

Every agent turn is a checkpoint. One click rolls files, transcript, and pending reviews back to before it ran.

▶ approval-gated

A live terminal, on your terms

Commands stream real stdout/stderr only after you approve; Stop kills the whole process tree — no orphaned servers.

background

Servers keep running

Long commands detach and keep streaming while the agent works on; it polls output, sniffs the port, tests against the live process.

transparent

Real token accounting

A live context meter with actual usage from the provider — click it for the full breakdown. Cost opacity is the enemy.

L02sketch builder

Drag a whole system onto the canvas.

Predefined agent topologies, ready to run — exactly as they appear on the real board. Drag the Design a Key-Value Store card onto the canvas and it lays down the eleven wired agents from the ByteByteGo chapter — requirements, architecture, storage engine, consistent hashing, replication + quorum, vector clocks, gossip + failover, integrate + server, tests, build + validate, and a review / run-guide. Or drop the Real-time Gaming Leaderboard card for a full Redis-sorted-set build. Set the goal at the top, hit Run, and watch the agents light up in order.

template palette

Drag a card onto the canvas to lay down a fully-configured agent topology.

Drop a template here

Drag a template card in — or just click it — to materialise the full topology.

L03core systems

Four subsystems, one editor.

reflex

Autocomplete

Debounced ghost text over every file, cancelled the instant you keep typing. Routed to a fast model per provider.

⌘⌥Icognition

Chat

Streaming chat that knows your project — pin files or let ripgrep + workspace symbols pull context automatically.

⌘⌥Erevision

Edit

Select code, state the change, review a true side-by-side diff. Keep or discard — nothing applies behind your back.

defaultautonomy

Agent

State the goal. It plans, edits across files, runs behind approval gates, and verifies before reporting.

np++ toolkit

Power-editing pack

Keystroke macros, line ops, column mode with incrementing numbers, encoding/EOL controls, big-file mode, hot-exit.

hack layer

Init script + packages

Code that runs at boot and hot-reloads on save; a package generator with a live dev loop; Atom / NPP / Sublime keymaps.

continuity

Sync + import

Built-in Settings Sync, notify-only updates, and one-click import from VS Code, VSCodium, or Cursor.

L04any model · your key

Fuse it to any intelligence — from your machine.

One registry, hundreds of models. Anthropic runs on a native adapter with prompt caching and full tool use; everything else flows through one OpenAI-compatible pipe. Requests travel from your machine to the provider — no relay, no logging, keys sealed in your keychain.

Anthropic

native · prompt caching + full tool use

OpenAI

o-series aware

OpenRouter

browse 400+ models live

Groq

fastest tokens on earth

Together
Fireworks
DeepSeek
xAI
Mistral
Ollama

local · fully offline

Custom

any OpenAI-compatible URL

+ hundreds of models through one registry — adding a provider is a data row, not a plugin

0

telemetry events phoned home

0

servers between you and your model

0+

providers · hundreds of models

0%

of AI runs on your own key

L05get atom++

Take the Atom home.

Free and open. Drag the .dmg to Applications, add a key — or run fully offline on local Ollama — and the orbits come alive.

Install in one drag

Download the .dmg, open it, drag Atom++ to Applications. Add your API key — or point it at local Ollama — and the editor starts editing with you.

free · MIT · no account required · all releases →

# on an early unsigned build, if macOS blocks first launch:
xattr -dr com.apple.quarantine "/Applications/Atom++.app"

Or build it yourself

The whole editor bootstraps from source in four commands — Code-OSS is fetched at a pinned tag and turned into Atom++.

$ ./scripts/bootstrap.sh  # clone + brand + patch
$ ./scripts/run-dev.sh  # launch from source
$ ./scripts/build-macos.sh  # produce Atom++.app
$ ./scripts/make-dmg.sh  # wrap into a .dmg

prereqs: Xcode CLT · Node 24 · Python 3.8+