hesper.

Every Mac you own, running your agents. One fast window to steer them.

Hesper turns all your Macs into one pool of compute for Claude Code and Codex agents. Start an agent on the laptop, the Mac mini under the desk or the Studio in the office, from any of them, and hand it to another Mac whenever you like. Every agent shows up as a live terminal tile on one wall, wherever it runs. The ones waiting for you line up in a single queue, one key away.

Every layer is trimmed for speed: a native AppKit app on libghostty, a small Go daemon per Mac, and a budget for every hop. Typing into an agent reaches the screen in about 3 ms, the wall holds 120 Hz with 16 live agents, and 30 idle agents cost the daemon practically nothing.

Hesper is named after Hesperos, the evening star: the first light after sunset. Your agents run quietly across your Macs; Hesper lights up the one that needs you.

Hesper is built on libghostty. It is an independent project and not affiliated with Ghostty.

The current release is v0.3.4, signed and notarized for macOS 14+ and installed with Homebrew. Every feature this README describes is in it. v0.3.4 lets agents name their own tiles after what they work on; v0.3.3 shows project bands whenever agents run in more than one project; v0.3.2 makes a wall tile active with one click (⌃⌥ + arrow from the keyboard); v0.3.1 fixes Claude Code's /voice inside Hesper. New since v0.2.0:

• Every action works with the mouse: buttons with key caps on tiles, sheets and popovers, right-click menus, close and back buttons
• A complete menu bar and a Dock menu, with items that disable when they do not apply
• Help ▸ Keyboard Shortcuts (⌘/) and rebinding in Settings ▸ Shortcuts
• VoiceOver and Full Keyboard Access support
• hesperctl open review and hesperctl open palette

See the v0.3.4, v0.3.3, v0.3.2, v0.3.1 and v0.3.0 release notes for the full list, including the six shortcuts that moved.

One Mac works without a relay. Multiple Macs currently require a self-hosted relay and device setup; Hesper does not provide a hosted cloud service. See More than one Mac.

One Mac runs out of cores, memory and battery long before you run out of tasks. Most people who run many agents have a second or third Mac sitting idle. Hesper uses all of them.

Start anywhere, run anywhere. Pick the machine when you start an agent (@mini in the composer, --machine mini on the command line), or give a project a default machine. A remote agent looks and behaves exactly like a local one: the machine is a badge on the tile, not a place you go to.

Move work to where there is room. When the laptop runs hot or its battery runs low, hand the busy agents to a Mac with room to spare (see Pick up anywhere).

Agents that use the whole pool. hesperctl, the Hesper skill and the MCP server let Claude and Codex start their own helpers on any of your Macs, wait for them and collect their results. Each agent controls only the agents it started, and only within depth and count limits.

Your machines, end to end encrypted. Macs on the same network talk directly; otherwise through a relay you run that forwards traffic it cannot read. Device keys live in the Secure Enclave, every device is approved per Mac, and opening a shell on another Mac needs Touch ID. Macs with agents at work stay awake (on battery only above 20 %).

Most agent tools tie a session to the machine it started on. In Hesper, work moves between your Macs as easily as it moves between windows.

Hand off a running agent. Continue on mini (⇧⌘M, ⌘K, or the tile's menu) moves an agent to another Mac. Hesper checkpoints its worktree (staged, unstaged and untracked work, never your own branches or index), sends only the changes over your encrypted link, recreates the worktree on the same branch there, and resumes the same Claude or Codex conversation. One ⌘Z moves it back. Fork on mini keeps the original and starts a twin.

Start anywhere, even without the folder. Start a task for a Mac that doesn't have the project yet, and Hesper brings it along: cloned from its remote with your uncommitted changes, or copied whole when there is no remote (scratch folders and plain folders included, build output left behind).

One history across all your Macs. ⌘Y searches every Claude and Codex session on every Mac. Resume one where it ran, continue it on another Mac, fork it, or hand a Claude session to Codex. When an agent closes, its last checkpoint stays with its session, so you can restore the work later.

Safe by design. Every piece of work lives in exactly one place at a time, and moving it is an explicit step you can undo. Nothing syncs your folders behind your back, so two agents never fight over the same files.

Hesper is meant to sit open all day next to dozens of busy agents, so every part is measured against a budget and kept lean:

How it stays fast:

• Native, no web view. AppKit and libghostty's GPU renderer, Swift 6 on the app side, Go on the daemon side. The real claude and codex TUIs run unchanged, so there is no chat UI to keep in sync.
• Draw only what is seen. Tiles are vsync-paced and draw only the rows they show; hidden and offscreen tiles render nothing. The focused agent skips vsync for the lowest typing latency.
• The daemon is the terminal. hesperd keeps one VT emulator per agent, so attaching, switching or reconnecting starts with an exact redraw instead of replaying output. Slow viewers get a fresh redraw instead of a growing buffer.
• Pushed, not polled. Claude's and Codex's own hooks tell the daemon when an agent works, waits or finishes, and every change is pushed to the app and to other Macs.
• Direct when possible. Macs on the same network skip the relay; a link moves to the direct path as soon as it is up.
• Background work stays in the background. Session history is indexed into SQLite FTS5 at background priority, and nothing runs for an agent that is quiet.

One queue, one key. ⌘J jumps to the next agent that needs you: approvals first, then questions, then errors. ⏎, A or N allows, always-allows or denies right on the tile. Tiles never jump around to get your attention.

Agents belong to the Mac, not the window. hesperd owns every agent's terminal. Close the app, restart it or reboot: agents come back with their session. You get tmux-style persistence without a terminal inside a terminal.

Fast to start. ⌘N opens a draft tile in the wall. Type the task, add @machine, #project, /profile or ~branch, and ⌘↩ starts it; ~branch gives the agent its own git worktree.

 ┌──────────────────────────── your Mac ─────────────────────────────┐
 │                                                                   │
 │  Hesper.app (AppKit + libghostty)        hesperctl (CLI)          │
 │        │  JSON-RPC over a 0600 local socket   │                   │
 │        ▼                                      ▼                   │
 │  hesperd (LaunchAgent) ── owns PTYs ── claude / codex / shell     │
 │        ▲                                      │                   │
 │        └──────────── hooks: state changes ────┘                   │
 └────────┬──────────────────────────────────────────────────────────┘
          │ outbound only, Noise IK end to end   ┌─── another Mac ───┐
          ├──────────► hesper-relay ────────────►│      hesperd      │
          └────────── direct TCP on the same LAN ►│                   │
                                                 └───────────────────┘

The full design, wire contract and "as built" notes are in docs/rebuild-contract.md; the security design is in docs/remote-shell-contract.md and relay/docs/.

• libghostty, embedded directly. A pinned, checksum-verified GhosttyKit.xcframework (no Zig toolchain needed). Only app/Sources/Hesper/Engine/ touches it. Each tile's command is hesperd attach <id>, so libghostty drives an ordinary PTY. Tiles stay vsync-paced; the focused agent renders without vsync for lower keystroke latency; hidden tiles render nothing.
• The daemon is the terminal. hesperd keeps its own VT emulator per agent, and every attach starts with an exact redraw of the screen. The emulator is tested against tmux as ground truth: the same bytes go into both and the screens must match. Slow viewers are resynced with a fresh redraw instead of buffering without bound.
• State from hooks, not screen scraping. The installer merges hooks into Claude Code's settings and keeps everyone else's. Codex agents get per-agent hook overrides plus a precomputed trust hash, so ~/.codex is never modified. hesperd hook returns within 250 ms and always exits 0, so a hook can never stall an agent.
• Resume you can trust. Agents are persisted atomically and resumed with claude --resume or codex resume only when the session verifiably exists. The original task is re-sent only if the agent never started, so "push the branch" never runs twice.
• A relay that cannot read or forge. Requests are signed with P-256 device keys (method rights, clock window, persisted nonce cache) and carried in a Noise_IK_25519_ChaChaPoly_BLAKE2s channel with pinned host keys. Terminal streams get per-direction keys; a tampered, dropped or reordered frame ends the stream. Pairing codes are compared on both screens, so a relay that swaps keys is caught.
• Checkpoint-based moves. hesperctl move ID --to mini checkpoints the agent's folder, packs the transcript and an incremental git bundle (the checkpoint carries the uncommitted and untracked work, never ignored files), sends it sealed, and resumes the agent on the target in a worktree on the same branch (cloning the project from its remote when the target lacks it), with a handover note; --fork keeps the original. A move that fails leaves the agent where it was. Checkpoints are also taken when a turn ends and at close (hesperctl checkpoint ID), kept 14 days.
• Shared history. Claude and Codex transcripts from every Mac are indexed into SQLite FTS5 (pure Go, no cgo) at background priority. Project data shared between Macs is a CRDT, so every Mac converges.
• Tested end to end. 275 Go test functions, 193 Swift Testing tests for the UI-free HesperCore module, about 300 in-app UI self-test checks run against a fake daemon and the real one, an in-app frame probe and perf harness, and an installer test suite that runs every scenario in a temporary HOME with stubs that refuse any system change.

Requires macOS 14 or later. Signed releases include the app, daemon, hesperctl and the Secure Enclave helper; Xcode and Go are only needed for source builds. Install Claude Code or Codex separately to run those agents.

The repository and signed release downloads are public. No GitHub account or token is needed to install.

With Homebrew installed:

brew tap derzierau/hesper https://github.com/derzierau/hesper.git
brew install --cask derzierau/hesper/hesper
/bin/sh /Applications/Hesper.app/Contents/Resources/hesper-setup.sh
open /Applications/Hesper.app

Homebrew selects the native Apple Silicon or Intel archive and installs the CLI tools on your PATH. Use the app path printed by Homebrew if you choose another --appdir. To update:

brew update
brew upgrade --cask derzierau/hesper/hesper
/bin/sh /Applications/Hesper.app/Contents/Resources/hesper-setup.sh

Run setup after each update to restart the daemon with the new binary.

The current tap lives in this repository, so brew tap above includes its HTTPS URL. A separate public derzierau/homebrew-hesper repository would allow the shorter brew tap derzierau/hesper command; see Homebrew's tap documentation.

Hesper is not yet in Homebrew's official catalog. Installation without adding a tap requires a separate submission and acceptance into homebrew/cask under its package acceptance policy.

Open GitHub Releases and download the archive for your Mac:

Each archive contains Hesper.app, including hesperd, hesperctl and hesper-keys in Contents/MacOS. Both architectures are Developer ID-signed, notarized and stapled. SHA256SUMS contains the archive checksums.

Extract the zip, move Hesper.app to /Applications, then run:

/bin/sh /Applications/Hesper.app/Contents/Resources/hesper-setup.sh
open /Applications/Hesper.app

The CLI is also available directly at /Applications/Hesper.app/Contents/MacOS/hesperctl.

After either binary installation, open Codex and run /hooks once to enable its trusted hooks. Setup installs the LaunchAgent, agent hooks and Hesper skill.

You need Xcode 26 (used from the command line) and Go 1.25 or later (brew install go).

git clone https://github.com/derzierau/hesper ~/projects/hesper
cd ~/projects/hesper
./install.sh --dry-run   # print every action, change nothing
./install.sh

The installer:

1. builds the relay tools and the app;
2. puts hesperd, hesperctl and hesper-keys inside Hesper.app and installs it to ~/Applications;
3. links hesperd and hesperctl into ~/.local/bin;
4. runs hesperd serve as a LaunchAgent;
5. merges Hesper's hooks into ~/.claude/settings.json and Codex (your other hooks are kept);
6. links the Hesper skill for Claude Code and Codex (see below);
7. lets hesperd through the macOS firewall.

Run it again after every pull; it only changes what differs, and anything it replaces is backed up first. See ./install.sh --help for --login-item, --skip-build, --no-firewall and the rest.

Codex only runs hooks it trusts: after installing, open Codex and run /hooks once.

Maintainer instructions for signing credentials, release automation and the Homebrew tap are in docs/releases.md.

Signing. Without configuration the app is signed ad-hoc, which works locally. SIGN_IDENTITY="Developer ID Application: …" ./install.sh signs with the hardened runtime, and NOTARY_PROFILE=<profile> also notarizes. macOS notifications need a Developer ID-signed build.

New agents start with each tool's permission prompts on (claude, codex); Hesper queues the prompts for you (⌘J). To run agents unattended, pick a profile per agent, per project or per kind:

Esc always goes to the agent, since Claude and Codex use it. Every key and menu item comes from one table in the app (⌘/ and Settings › Shortcuts list it by surface). The defaults above can be changed in Settings › Shortcuts: click a shortcut, press the new one (⌫ removes it, Reset brings the default back); a chord macOS keeps or another command already has is refused with the reason. Every row, card and tile is a button to VoiceOver and Full Keyboard Access, and ⇥ walks the window. The menu bar item shows the counts and the queue, and a notification arrives when an agent starts waiting for you, or when its finished work is ready to review.

hesperctl does the same from a terminal:

hesperctl ls                                  # every agent, with its state
hesperctl new --project ~/projects/app "Fix the login redirect"
hesperctl new --machine mini --worktree --project ~/projects/app "Add the export"
hesperctl attach fix-the-login-redirect       # Ctrl-] detaches
hesperctl send ID "yes, go ahead" ; hesperctl approve ID ; hesperctl deny ID
hesperctl stop ID ; hesperctl resume ID ; hesperctl rm ID ; hesperctl move ID --to mini
hesperctl open review --agent ID ; hesperctl open palette --query TEXT   # the app's sheets
hesperctl rename ID "Release notes" ; hesperctl rename ID --auto   # your name stays; --auto hands it back

Agents are addressed by id (L/a7f3k2), local id or unique name. Claude and Codex agents name their own tile after what they work on, in your language (hesperctl rename self), and rename it when the work moves on; a name you give stays.

hesperctl mcp serves the same commands as MCP tools on stdio, so Claude Code and Codex can start, watch and answer agents themselves (hesper_agents_new, hesper_agents_wait, hesper_history_search, …, plus the resources hesper://agents, hesper://needs-you and hesper://reference). ./install.sh --mcp registers it with both (claude mcp add --scope user hesper -- ~/.local/bin/hesperctl mcp, and [mcp_servers.hesper] in ~/.codex/config.toml). Inside a Hesper agent the agent-tree rules apply to its calls: what it starts are its children.

skills/hesper is an Agent Skill that teaches Claude Code and Codex to use hesperctl: see what every agent is doing and which ones need you, fan work out to parallel agents in worktrees and collect their results, delegate and wait (hesperctl new --wait), drive or answer an agent, and find and resume past sessions. It keeps to Hesper's agent policy (an agent steers only the agents it started) and never approves anything on your behalf unless you ask. The installer links it into ~/.claude/skills/hesper and ~/.agents/skills/hesper (Codex); --no-skills leaves it out. Then just ask, e.g. "what are my agents doing?" or "split this into three agents in worktrees and report back". hesperctl reference is the full CLI it reads.

Every Mac runs its own hesperd. One Mac needs nothing more. To use agents across Macs you need a relay: a small Go server you deploy yourself (Docker, behind Caddy or nginx; there is no built-in or shared one). The Macs sign in to it and find each other through it; on the same network they then talk directly, otherwise through the relay. relay/docs/deployment.md covers deploying it; then point each Mac at it and sign in:

./install.sh --relay https://relay.example.com   # your relay's address
hesperctl login --role host --name "Mac mini" --out ~/.local/state/hesper/host.credentials.json
hesperctl login --role controller --name "Mac mini" --out ~/.local/state/hesper/controller.credentials.json

relay/README.md covers the sign-in, where the relay address comes from, and approving devices with hesperctl pair-host and hesperctl approve.

Opening a shell on another Mac is off by default. Install with --allow-shell on the Mac that should accept it; the request then needs Touch ID on the Mac that sends it.

Set HESPER_STATE_DIR to use another state directory.

cd relay && go vet ./... && go test -race ./... && make build   # hesperd, hesperctl, relay, keys
make -C app test-unit build                                       # HesperCore tests, then Hesper.app
make -C app run                                                   # the app against a fake daemon
make -C app test                                                  # plus UI self-tests (needs a GPU)
scripts/test-install.sh                                           # installer, dry runs in a temp HOME

Contributions are welcome; see CONTRIBUTING.md. Report security issues privately as described in SECURITY.md.

Hesper is young and used daily by its author. Signed releases are available through Homebrew and GitHub Releases; the current one is v0.3.4.

Supported agents are Claude Code, Codex and a plain shell.

Default profiles now keep permission prompts on. Earlier versions started new Claude agents with claude-auto-rc (auto mode) and new Codex agents with codex-full (approvals and sandbox off) unless you configured otherwise. Now the defaults are plain claude and codex. Nothing in your ~/.config/hesper is rewritten:

• profiles and defaults you set in profiles.json / settings.json keep working as before;

• the old names claude-bypass and codex-full still work wherever they are saved (running agents, project and per-folder defaults, drafts, scripts) and mean claude-unattended and codex-unattended;

• if you relied on the old built-in defaults, new agents now ask before acting. To keep the old behaviour, put this in ~/.config/hesper/settings.json and restart hesperd (launchctl kickstart -k gui/$(id -u)/de.olezierau.hesperd):

  {"defaults": {"kinds": {"claude": "claude-auto-rc", "codex": "codex-unattended"}}}

Hesper was formerly called Ghosty. On a Mac that ran it, install.sh stops the old app and LaunchAgent, moves state and config to their hesper names, rewrites saved paths and replaces the hooks. Agents, history, device keys and relay enrollments carry over, and every agent resumes under hesperd. A few wire names keep the old spelling for compatibility.

MIT. Geist and Geist Mono are bundled under the SIL Open Font License. Third-party licenses are listed in THIRD_PARTY_NOTICES.md. A build from source with the default prebuilt libghostty includes GNU libintl (LGPL-2.1); release builds do not. docs/licensing.md explains both and how to build either way.