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.

What is agentglass?

agentglass is a desktop application that gives you a single window to monitor and interact with the coding‑agents you already run on your machine (Claude Code, Codex, Gemini CLI, OpenCode, etc.). It hooks into the tmux sessions and Git worktrees those agents use, then surfaces their terminals, pull‑request diffs, browser interactions, approvals, Docker containers, logs, and resource usage—all without you having to alt‑tab between many separate tools.

Core ideas

Idea What it means
One unified workspace All agent‑related UI (tmux panes, git diff viewer, PR panel, browser, approvals) lives inside one Electron/React window.
Attach, don’t replace agentglass starts the same CLI you already use; it merely attaches to the tmux session so the agent keeps running even if you close the UI.
Local‑first & privacy‑first Everything is stored in a local SQLite file. No cloud account, no telemetry. The only outbound actions are an optional webhook you configure yourself and an “Explain” button that can send selected diffs to a model (both off by default).
Extensible via plugins Plugins run as separate processes with scoped, revocable tokens and can be installed from an in‑app catalogue.
Mobile companion An Android app can pair with the desktop instance (QR‑code scan) to let you view approvals, PRs, and terminal output on the go.

Main features

  • Agent management – launch agents from tasks or GitHub issues; each creates a git worktree, branch, and tmux pane.
  • Real tmux terminals – view and control actual tmux panes; click a make target or npm script to run it.
  • Lantern status bar – red/amber indicators show agents waiting on permission gates or finished work awaiting your input.
  • Diff viewer – word‑level changes grouped by agent/worktree; you can stage, commit, stash, or resolve conflicts directly.
  • Pull‑request panel – see checks, conversation, files, and write an inline review without opening a browser.
  • Browser panel – agents can drive a built‑in Chromium view; the browser is already signed‑in with your credentials.
  • System view – Docker containers, logs, listening ports, CPU/memory, and cost tracking per session.
  • Android companion – same panes and approvals on your phone via a QR‑code pairing.
  • Plugin system – add new panels or integrations; plugins are sandboxed with fine‑grained tokens.

Installing & running

  1. Download the appropriate binary from the latest release (AppImage, .deb, .dmg, .exe, or .apk). No separate server process is needed.
  2. macOS note – the unsigned build triggers Gatekeeper; run xattr -dr com.apple.quarantine /Applications/agentglass.app once.
  3. From source (requires Bun, Python 3, git, tmux, Claude Code CLI):
    git clone https://github.com/SirAllap/agentglass && cd agentglass
bun install
AGENTGLASS_STATE_DIR=~/.local/state/agentglass-dev bun run dev   # UI at http://localhost:4000
python3 hooks/install_hooks.py   # registers Claude Code hook
  4. Follow docs/INSTALL.md for platform‑specific prerequisites (git, tmux, GitHub CLI, Docker, etc.).

Platforms supported

OS Status Distribution format
Linux ✅ Working AppImage, .deb
macOS ✅ Working (unsigned) .dmg (Intel & Apple‑silicon)
Windows 🧪 Unverified on real hardware (build passes CI) .exe
Android ✅ Working .apk (paired companion app)

Security & privacy

  • Data stays local – a single SQLite file on your disk holds all state.
  • No default outbound traffic – only two opt‑in features can leave the machine:
    1. A user‑configured webhook.
    2. The “Explain” button, which sends selected diffs to a language model.
  • Token‑scoped plugins – each plugin runs in its own process with a revocable token limiting what it can access.
  • See SECURITY.md for the full trust model, retention policy, and how to report vulnerabilities.

Extensibility

  • Write a plugin by following docs/EXTENDING.md and PLUGINS.md – define the HTTP API you need, request a scoped token, and ship it via the in‑app catalogue.
  • The architecture is a Bun/SQLite backend, a React/Vite UI, Electron for the desktop shell, and a small Python forwarder for Claude Code hooks.

Roadmap highlights

  • Now: Lantern‑based verdict strip, macOS signing/notarization, collision warnings for parallel agents.
  • Next: Rule‑based tool gating, per‑project gate policies, MCP‑based cockpit queries, Windows hardware verification.
  • Later: In‑place diff review with single‑prompt submission, API‑panel for fleet endpoints, decision‑log mining, voice input.

Community & contribution

  • Discussions, issue tracker, and a “help‑wanted” label are active on GitHub.
  • Contributions are encouraged; the project aims to stay lightweight (Bun, SQLite, React, Electron, React‑Native, stdlib Python).
  • License: MIT (2026) with a separate CC‑BY‑4.0 attribution for bundled artwork.

In short: agentglass is a privacy‑first, locally‑run UI that lets developers keep all of their AI coding‑agent workflows—terminals, PR reviews, diffs, browser actions, and approvals—in a single, extensible window.

Related

  • Project
    LucieEveille/kiwi-mem

    kiwi‑mem is a self‑hosted FastAPI gateway that adds human‑like long‑term memory to LLM chats. It stores facts in PostgreSQL + pgvector, assigns each a heat that decays over time, merges related pieces during a nightly “Dream” cycle, and compresses old chats into hierarchical calendar summaries. The system works with any OpenAI‑compatible or Anthropic native API, provides a web admin panel, and runs via a single Docker‑Compose command.

  • Project
    PinelliaChill/Operant

    Operant is a local, extensible AI‑agent platform that lets a single user run multiple LLMs in coordinated workflows (planning, coding, reviewing, etc.). It offers safety‑first approvals, a versioned memory system for re‑using experience, plugin‑based tool extensions, and several interfaces (CLI, web UI, TUI, desktop). All state is stored locally and model calls go to any OpenAI‑compatible API you configure. The project is in beta (macOS arm64 & Ubuntu 24.04 validated) and is licensed for non‑commercial use only.

  • Project
    hermes-hq/hermes-ide

    Hermes IDE is a cross‑platform desktop app that lets you run multiple code‑generation agents (Claude, Codex, Copilot, etc.) side‑by‑side, each in its own real terminal and Git worktree. It provides inbox notifications, per‑turn diffs, a Review Desk for merging, and optional structured Claude‑view UI, all built with Tauri, React and Rust.

  • Project
    huisezhiyin/sdd-riper

    SDD‑RIPER Light is a lightweight, repo‑native framework that adds a control plane around LLM‑based coding agents. It defines a clear loop—restating goals, creating a minimal spec, awaiting human approval, executing, validating with evidence, and syncing results back—so agents can modify code safely, auditable and recoverable. Four modular “skills” (light, strict, codemap, new‑chat‑ready) cover daily work, high‑risk tasks, unfamiliar codebases, and hand‑offs.