AgentGlass Review 2026 — AI Coding Agent Observatory That Puts Every Agent on One Screen

Overview

AgentGlass (stylized agentglass) is an open-source observability cockpit that brings every AI coding agent on your machine onto one screen. Launched in mid-July 2026 by SirAllap, it has already amassed 273+ GitHub stars and is being actively developed with 390+ commits from both the maintainer and Claude Code contributions.

The core promise is simple but powerful: stop context-switching between Claude Code terminals, OpenAI Codex sessions, Gemini CLI runs, and Bespoke agent dashboards. Point agentglass at all of them, and watch live cost, tokens, tool calls, errors, and session lifecycles in real time — from one cockpit.

What Makes AgentGlass Different

Most agent observability tools are either per-provider dashboards (like Anthropic’s own console) or lightweight event streams that forget everything on refresh. AgentGlass sits as a persistent visibility layer that:

Connects via Claude Code hooks (natively) or OpenTelemetry GenAI exporters for OpenAI Codex, Gemini CLI, Bedrock, LangChain, LiteLLM, and others
Scans ~/.claude/projects transcripts so historical sessions appear instantly, even before agents start streaming
Persists everything across reloads — not a “live only” dashboard
Filters by provider, project, or scope — zoom into one repo or see the whole machine

The data model includes cost tracking (per-provider pricing), tool-latency percentiles, error timelines, session lifecycles, and a sweeping agent radar that visualizes context window usage — a blip near the edge of the radar is about to compact.

The Workspace: More Than a Dashboard

What elevates AgentGlass beyond a monitoring tool is its built-in workspace — six keyboard-driven panels that let you act on what the fleet produces without leaving the cockpit:

Diff & Review

A full Shiki-syntax-highlighted diff viewer with word-level intra-line diffs, split/unified modes, and one-click “Explain” (local Claude walkthrough of the change set). Review and commit directly from the diff view.

Source Control (lazygit-style)

Stage, hunk-stage, commit, branch, stash, push/pull — all live on any repo the fleet has touched, all write-gated (read-only until you opt in). Interactive hunk staging, conflict resolution with per-side selection, and merge undo.

Pull Requests

Review PRs without opening a browser: overview, conversation, commits, files, checks, and review composer. Saved views, facet filtering (author, label, checks, etc.), and merge/squash/close actions.

Docker (lazydocker-style)

Containers grouped by compose project with live CPU/memory stats, log viewer (colored by level, pinned to bottom), and start/stop/restart actions. Exec into a container directly.

Real Terminal

A genuine PTY shell (your login shell) per repo/worktree over a WebSocket. Persistent sessions survive panel closes. Plus a commands menu that discovers Makefile targets and package.json scripts from the whole project tree.

Claude Chat

Drive local Claude Code sessions from the browser. Pick model + permission mode, stream replies, resume existing sessions. Sessions started here appear in the fleet like any other agent.

Phone Companion

AgentGlass ships a separate mobile interface (not a responsive stylesheet — a different React tree) designed for triage, not terminal work. The home screen is a queue of decision cards:

Card Type	What It Means
Blocked · waiting on you	An agent gate — approve or deny a dangerous command
Container down	A container exited non-zero — tail logs or restart
CI went red	A failing check on your PR — open log or re-run
Ready to merge	Approved, checks green, nothing blocking
Review requested	Somebody asked for your eyes — opens PR with diff
Stopped · wants direction	A session that ended its turn 4min–12h ago

The phone app has only three tabs: Now (the queue), Chats (conversations that persist across sessions), and Repos (changes, PRs, containers). It connects via QR code on your LAN and uses a 32-character secret.

Pricing

AgentGlass is completely free and open source under the MIT license. The server (Bun + SQLite) and desktop app (Electron) are self-hosted on your machine — no cloud dependency, no API keys needed beyond what your agents already use.

Community & Activity

GitHub Stars: 273+ (as of July 28, 2026)
Commits: 390+ in ~2 weeks since public release
Issues: 40 open
Pull Requests: 2 open
License: MIT
Tech Stack: Bun, SQLite, React, Vite, Motion, Shiki
Desktop: Electron (macOS, Windows, Linux)
Trendshift: Ranked on daily trending for TypeScript

Verdict

AgentGlass fills a genuine gap in the AI coding agent ecosystem. As developers increasingly run multiple agents (Claude Code, Codex, Gemini CLI) across multiple projects and providers, the need for a unified observability layer becomes urgent.

What	Score
Ease of Use	8/10 — The cockpit is polished but dense; getting started requires understanding hooks or OTel
Features	9/10 — Workspace panels (diff, git, Docker, terminal, PRs, chat) go far beyond any other dashboard
Value	9/10 — Free, MIT, self-hosted, no lock-in
Performance	8/10 — Bun+SQLite backend is snappy, but the Electron desktop app is memory-heavy
Ecosystem	7/10 — New project; Claude Code hooks are the most mature path, OTel support is still expanding

Overall: 8.2/10 — Silver+

For developers who want to understand what their AI agents are actually doing, costing, and breaking — and act on it from one place — AgentGlass is the best option available today. The workspace feature alone puts it ahead of every other agent observability tool on the market.

How to Get Started

# Install via npm
npx agentglass

# Or clone and run from source
git clone https://github.com/SirAllap/agentglass.git
cd agentglass
bun install
bun run dev

Then configure Claude Code hooks or set up OpenTelemetry exporters, and open http://localhost:5173 to see every agent on your machine.