SirAllap/agentglass

🛰 Every AI coding agent on your machine, on one screen — live cost, tokens and tool calls across every provider, and a hold on anything dangerous until you say go. From your desk or your phone

Summary

The agentglass repository is a local desktop workspace for supervising AI coding agents without moving between terminals, browser tabs, and approval windows. It attaches to existing tmux sessions and repositories, combining Claude Code, Codex, Gemini CLI, and OpenCode sessions with diffs, pull requests, a browser, and approval gates, plus an Android companion for remote approvals. The TypeScript project uses a Bun/SQLite server, React/Vite UI, Electron shell, React Native app, and Python hook forwarder, with state kept in a local SQLite file and no account, cloud service, or telemetry.

What it is useful for

Use it to coordinate parallel agent work without replacing the CLIs. A task or GitHub issue can become a worktree, branch, and running agent; an orchestrator can delegate to workers under a policy you control, with push and merge out of their reach. Agents stay in tmux and keep running if you close agentglass.

The Lantern handles approval triage: red means an agent is blocked on a permission gate, and amber means it has finished and is waiting for input. A gated tool call remains held until you allow it from the desktop or phone, with the queue stored on disk so a crash cannot silently allow it.

For review and operations, the workspace groups word-level diffs by agent and worktree, supports staging hunks, commits, branches, stashes, and conflict resolution, and brings pull-request checks and conversations into the app. It also exposes tmux panes, project scripts, Docker resources and logs, listening ports, CPU and memory use, and session cost and time.

The Android companion can show terminals, approvals, pull requests, and changes at the scope you grant during pairing.

Important boundaries: agents run as your user, so this is a lock, not a sandbox. Windows is not verified on physical hardware, and macOS builds are unsigned. Worktrees begin with only Git-tracked files, so .worktreeinclude is needed for ignored paths such as .env or local settings. Webhooks and Explain are off by default.

How engineers can use it

Install a platform build from Releases, or follow the README’s source setup:

git clone https://github.com/SirAllap/agentglass && cd agentglass
bun install
AGENTGLASS_STATE_DIR=~/.local/state/agentglass-dev bun run dev
python3 hooks/install_hooks.py  # so Claude Code reports to it

This runs the source server at http://localhost:4000 with its own database. The base requirements are git, the Claude Code CLI, and Python 3; tmux, the logged-in GitHub CLI, and Docker unlock specific features, and Settings ▸ Requirements checks them locally.

A practical first session is to open an existing tmux agent, launch a task from a GitHub issue, inspect its worktree diff, and resolve any red Lantern gate. Check docs/INSTALL.md for platform caveats and SECURITY.md before enabling outbound features.