agentglass

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.

🛰 What it does

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.

📸 Screen by screen

And all of it moving:

Every view, and what each is for: docs/WORKSPACE.md.

💻 Platforms

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

⬇️ Install

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 it

Full instructions, every platform's caveats and the requirements it expects to find: docs/INSTALL.md.

🔒 Local, and it stays that way

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.

❓ FAQ

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.

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.

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.

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.

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.

📚 Documentation

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

🧩 Two things worth knowing about

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.

🗺 Roadmap

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.

💬 Community

🤝 Contributing

Issues and PRs welcome — see CONTRIBUTING.md. Small, fast and