agentglass doesn't replace your agents. It attaches to the tmux sessions and repos already open on your machine, and puts every tool call, dollar and dangerous command on one screen — at your desk or in your pocket.
Every agent on one screen. Claude Code, Codex, Gemini CLI and OpenCode, in the tmux sessions and repositories already open on your machine. Every tool call as it happens, what each is costing, where the time goes, which session is stuck in a loop, what files it touched. agentglass does not launch them and does not replace them — it attaches to what is already running.
It tells you who needs you. The Lantern leads with one question and answers it: red for an agent stopped on a permission gate, amber for one that finished a turn and is waiting for you to type. Every agent is a card — model, branch, what it is doing this second, cost, turns, errors, and how long before its prompt cache goes cold. A watch re-reads it every few minutes and sends one notification when somebody has been left hanging.
Nothing dangerous runs unwatched. A tool call you decided to gate is held until you allow it, from the desk or from your phone. The queue is on disk, so a crash cannot silently auto-allow.
It is a place to work, not a dashboard. Read the diff the agent just wrote, review a pull request through to a verdict, run git and docker, drop into a real terminal with real tmux panes, drive a browser your agents can use.
 The cockpit Every session, what it costs, where the time goes, and which one needs you |
 Pull requests Review one to a verdict without opening a browser |
 Diff review Read what the agent just wrote, hunk by hunk |
 Terminal Real tmux panes, and the keys a phone does not have |
 Tasks A GitHub issue to a cut worktree in one press |
 In your pocket The machine's panes, a checkout's changes, and what this device was granted |
And all of it moving:
Every view, and what each is for: docs/WORKSPACE.md.
| Platform | Status | You want | Worth knowing |
|---|---|---|---|
| Linux | ✅ Working | .AppImage · .deb |
— |
| macOS | ✅ Working, unsigned | .dmg — Apple silicon & Intel |
One xattr command on first launch, below |
| Windows | 🧪 Not verified on real metal | .exe |
Builds and passes CI; a confirmation on hardware is wanted |
| Android | ✅ Working | .apk, signed |
Companion app, paired by a code you scan |
A desktop app with its own server inside it: nothing to run in a terminal, no port to open in a browser. Take the build for your platform from Releases and launch it.
On macOS the build is not signed yet, so Gatekeeper calls it damaged. It is
not: xattr -dr com.apple.quarantine /Applications/agentglass.app, once.
Or run it from source:
git clone https://github.com/SirAllap/agentglass && cd agentglass
bun install && bun run dev # http://localhost:4000
python3 hooks/install_hooks.py # so Claude Code reports to itFull instructions, every platform's caveats and the requirements it expects to find: docs/INSTALL.md.
Your machine holds everything: a SQLite file on your own disk, no account, no cloud, nothing phoned home. The phone reaches the desk over your own network, paired by a code you scan, at a scope you choose — read, answer, or full.
Two paths leave the machine and both are opt-in and off by default: a webhook you configure yourself, and the Explain button, which sends the hunks you point at to a model. Read SECURITY.md before you install: what is stored, what is exposed, what each token can do, and every switch that turns a capability off.
Does it replace Claude Code, or my agent?
No. It attaches to the tmux sessions and repositories already open on your machine. It does not launch your agents and it does not proxy them — close agentglass and every agent keeps running.
Does anything leave my machine?
No account, no cloud, no telemetry. Everything lives in a SQLite file on your own disk. Two paths can leave and both are off by default: a webhook you configure yourself, and the Explain button, which sends the hunks you point at to a model. SECURITY.md lists every switch.
What do I actually need installed?
Needed: git, the Claude Code CLI, and Python 3. Per feature: tmux (chats as live panes, your tmux windows as tabs, theme sync), the GitHub CLI (the pull-requests panel — and logged in, which is the step people miss), Docker (containers, images, volumes, logs).
Settings ▸ Requirements checks all of it on your machine and says what stands down without each one.
Why does macOS say the app is damaged?
It is not damaged — the build is not signed yet, and Gatekeeper says that about
anything unsigned. Once: xattr -dr com.apple.quarantine /Applications/agentglass.app.
Signing and notarization are the first item on the roadmap.
Which agents, and which providers?
Agents: Claude Code, Codex, Gemini CLI and OpenCode. Providers, for pointing Claude Code at something else: Kimi, OpenAI, Gemini and Bedrock — see docs/INSTALL.md.
| INSTALL.md | Installing, updating, requirements, the desktop app, the control plane, and running against any provider — Kimi, OpenAI, Gemini, Bedrock |
| WORKSPACE.md | Every view and what it is for: the rail, the Clone, the Lantern, the terminal, keyboard shortcuts, themes |
| CONFIG.md | Every environment variable the code reads, the whole HTTP API, and the architecture |
| SECURITY.md | The trust model, what each surface can reach, retention, and how to report a vulnerability |
| EXTENDING.md | Driving agentglass from your own harness, and writing a plugin |
| PLUGINS.md | What a plugin is, what it may ask for, and how to publish one |
| CHANGELOG.md | What each release changed |
Worktrees start with what git leaves out. git worktree add copies the
tracked tree and nothing else — no .env, no local settings — and a new
checkout does not start without them. A .worktreeinclude at the repository
root names the ignored paths every worktree agentglass cuts should carry in.
The terminal has two keyboard tricks. Ctrl+Shift+Space letters the paths,
links, hashes and ids on a pane's screen so one key pastes one back. And
selecting something in a pane offers to ask the agent in that pane about it.
Themes, not dates. The living version is the issue tracker; the
help wanted
label is where to start.
Now
- Signing and notarization for the macOS build, so Gatekeeper stops calling it damaged
- Warn when parallel agents collide on shared runtime the diff cannot see — #118
- Say when two live sessions share one working tree, instead of attributing the changes to whichever asked last — #117
Next
- A gate that can hold by rule — a tool allowlist beside the spend threshold that already works — #109
- Per-project gate policies and hook profiles — #14
- Let an agent query the cockpit over MCP: what is running, what it costs, what is held — #296
- Keep model prices fresh without hand-editing the table — #9
- Verify the Windows build on real hardware — #231, #195
Later / exploring
- Review a local diff in place and send the whole review as one prompt — #294
- An API panel to exercise the endpoints the fleet is building — #170
- Tasks per project, and a decision log mined from transcripts — #12, #13
Shipped so far, newest first: CHANGELOG.md and the releases.
Issues and PRs welcome — see CONTRIBUTING.md. Small, fast and
dependency-light on purpose: a Bun/SQLite server, a React/Vite UI, an Electron
desktop shell, a React Native phone app, and a stdlib-only Python hook
forwarder. To adapt it to another agent or harness without forking, start at
docs/EXTENDING.md.
Built by @SirAllap (David Pallares). Original work — not a fork. Not affiliated with or endorsed by Anthropic; "Claude" and "Claude Code" are trademarks of Anthropic.
MIT © 2026 David Pallares — see LICENSE.
One bundled set of artwork is not MIT: the portrait layers the Understudy draws are CC BY 4.0, which asks for attribution rather than permission. NOTICE.md is where that is discharged, and it is the only reason this repository has a NOTICE at all.