A board the agents share
Tasks with the files they own and what they depend on, claimed atomically over MCP. Hangar never runs two tasks at once whose files overlap.
Hangar runs the Claude Code and Codex (beta) already on your machine, side by side, as many sessions as you need, with a shared board, a shared memory and an orchestrator that hands out the work. Either one is enough. Talk to it in a chat, or watch every terminal at once.
Runs locally. No account needed, no telemetry.
Chat view Experimental
Switch a workspace to a chat with its orchestrator, Claude Code or Codex. The agent keeps running in the panes behind it, and the panes come back as they were.
Turn it on in Settings › Experiments.
Orchestrator, workers, deliveries
Pin a Claude Code or Codex pane as orchestrator: it splits the work into board tasks and starts a worker for each, of either agent, on a branch of its own in its own git worktree. With Claude Code, by default Opus plans, Sonnet works and Sonnet reviews.
Tasks with the files they own and what they depend on, claimed atomically over MCP. Hangar never runs two tasks at once whose files overlap.
A finished worker hands in its branch. Hangar merges it in a scratch worktree, runs your check command and has a reviewer agent read the diff. A step that fails goes back to the worker.
main in a scratch worktree
Your checkout is left alone until the fast-forward.
pnpm test passed
The check command you set, run by your shell.
main fast-forwarded
Local only. Nothing is pushed.
An illustration of the stages, with the task in these pictures.
Merges are local fast-forwards. Pushing, pull requests and releases stay your call.
Deliveries that wait together share one run of the check. If it fails, Hangar finds the one that breaks it and sends back only that one.
Where its delivery stands: queued, merging, checking, reviewing, landing, then how it ended.
The orchestrator rests while workers run. A delivery, a question or a worker that stops wakes it at once; the rest comes in a digest, every five minutes at most.
A worker busy with a long build or test is not reported idle, and one that is idle is reported once, not over and over.
Permissions, the check command and the branch work lands on are set per workspace and enforced by Hangar.
Hangar Mobile
See what waits for you in every project, answer your agents' questions with a tap, allow or deny a permission. Hangar serves the app itself, over your own Tailscale network: no app store, no account, no server of ours.
Hangar Mobile running on sample projects.
Hangar Mobile listens on 127.0.0.1 only: Tailscale,
a free app that links your own devices, is the only way in, over HTTPS. Each phone pairs with
a one-time code. Notifications reach it even with the app closed, sent by your PC.
Kits & previews
Over MCP, an agent opens images, code, Markdown, PDFs, web pages, videos and 3D models next to its pane, then renders snapshots to check its own work. Kits bring skills, MCP servers and tabs of their own.
glTF, GLB, OBJ, STL, PLY and FBX. A model reloads when its file changes, and glTF animations play.
A kit's tab is an HTML page, or a native one in Rust compiled to WebAssembly that Hangar draws in your theme.
They restyle the whole window, not only the terminals. Graphite is the default, where colour only marks a state.
Try one on this page
Experiments Off by default
Under Settings › Experiments. Unfinished work: they can change, break or disappear in any update.
Add a Linux machine you reach over SSH: Hangar installs a headless hangar-node,
starts it and keeps a tunnel to it. A workspace can live there, its terminals, agents
and board included, and the window shows it like a local one. It keeps working there with
your computer off.
Or send one worker to the node for a heavy build or a GPU box: its branch comes back to be checked, reviewed and landed here.
A workspace as a conversation with its orchestrator, its tasks beside it. See above.
Each pane in a systemd scope of its own: when memory runs out, the system stops the pane that uses it, a build say, not Hangar and every agent.
Performance
No Electron: Hangar draws its own window, and each terminal reads its output on a thread of its own. Measured, not estimated.
2.9%
Hangar's CPU with eight agents printing, chat view
was 8.1 %
3.2s
Eight sessions all back after a relaunch
was 9.1 s
0.38s
A workspace in the chat view, ready
was 0.66 s
79MB
The executable, with whole-program optimization
was 119 MB
249MiB
16 terminals printing non-stop for 30 s
1.12.1: 8.1 GiB
2.5ms
Echo of a key in a pane while 15 others print
1.12.1: no echo within a minute
0.57s
Cold start, four prompts readable
1.12.1: 0.96 s
Where they come from: 2.2's, in the changelog · how 2.0 against 1.12.1 was measured
Notes every agent can search and write, across workspaces: conventions, decisions, traps.
Either one is enough. An orchestrator of either agent hands tasks to workers of either agent, and each workspace keeps an account per agent.
An agent hands a sub-task to the other and gets the result back. What Codex draws shows in the chat, on your phone and on Discord.
Claude Code, Codex or any shell: PowerShell, cmd, WSL, zsh, bash, fish, Nushell.
Reopen the window and every agent pane resumes its conversation, the ones on screen first.
After 10 minutes untouched and out of sight, a workspace stops its agents to free memory, and resumes them when you return.
Tokens, active time and an API-equivalent cost by project, model and agent, read from your agents' own transcripts.
Hold a key and speak: whisper.cpp transcribes on your machine. Off until you turn it on.
One instruction to every pane; agent panes can hold it until their agent is free.
One agent that follows every workspace: ask where they stand, or pass an instruction to an orchestrator.
When neither Claude Code nor Codex is there, one click installs either with its maker's official installer; then you sign in.
Images, PDFs, code, videos and 3D models from anywhere on the computer, not only the project. A viewer loads only what was shown.
Experimental: a server of your own as a remote, a channel per workspace and a thread per worker, with the files agents share. ⚡ sends a waiting message at once.
Sign in with GitHub to see the statistics of all your machines together. Off by default; what it stores.
Install it once: it offers each new version and installs it in one click.
64-bit installer, for you alone, no administrator rights.
Download .exeNot code-signed yet: SmartScreen asks once. What to do
2.2.0, the latest for the Mac for now. Universal, macOS 10.15 or later.
Download 2.2.0 .dmgNew macOS builds are on their way. Details
x86-64. The AppImage runs on distributions from 2022 on (glibc 2.35).
Download .AppImageYou also need Claude Code or Codex (beta), or both: Hangar installs them if they are missing, and you sign in once. Every file of the latest release
Nothing at the moment, and no account is needed. The agents cost what they already cost you, on your own plan or API key. An orchestrator, its workers and a reviewer run at once, so they use a plan's limits faster than a single agent.
Yes. Hangar starts the claude and the codex you installed, with the
logins they already have, and never asks for your credentials. The navigator shows your plan's
five-hour and weekly usage.
Hangar drives Codex through the protocol OpenAI publishes for it, and was written from that protocol's schema, source and docs rather than from recorded sessions. Until Codex users confirm it, it stays beta. Something wrong in a Codex pane? Report it. What works with which agent: the Codex guide.
Not through Hangar: no telemetry, and your code, prompts and transcripts never reach a server
of ours. The board lives in your project (.hangar/board.json); memory, settings
and statistics stay in the app's data folder. Hangar goes online to check GitHub for updates,
for the dictation model or Groq if you turn them on, for Hangar Mobile's notifications and
Discord once you enable them, and for the optional Hangar account once you sign in.
No. It is off by default, and Hangar works the same without one: it never contacts the account server until you sign in. Sign in with GitHub, in Settings › Hangar account, to link your machines; turn on the statistics sync, a switch of its own, and the Statistics page shows all of them. It never receives a file path, a prompt or a line of code. What it stores, and how to delete it.
No. Workers commit on branches of their own, and a delivery lands on yours as a local fast-forward. Hangar never pushes, opens a pull request, tags or deploys, and its agents are told not to either.
It is not code-signed yet, so SmartScreen warns the first time: choose More info, then Run anyway.
Releases after 2.2.0 are built for Windows and Linux only. 2.2.0's universal disk image, signed and notarised, is the latest one for the Mac for now, and there is none to update it to yet. New macOS builds are on their way.
Hangar open on a PC that stays on, and Tailscale on the PC and the phone, on the same account. On an iPhone, notifications need iOS 16.4 or later and the app opened from the Home Screen. Merging a delivery is still done on the PC.
On the Discord. Bugs and requests go to GitHub issues.