v0.4.0

The cockpit stops lying to you about what it can see.

Every fix in this release is the same shape: something reported healthy, or reported nothing at all, while the truth was elsewhere. The desktop shell adopted whatever answered on :4000. The sync button refused the one merge that would work. A restart auto-approved decisions a human was in the middle of making. An update sat published and never said so.

Desktop and install

• The shell no longer adopts a foreign server. /health names itself, and when another app holds :4000 the sidecar moves to the next free port instead of pointing the whole cockpit at a stranger (#127).
• Rebuilding stops the running app properly rather than signalling its whole process tree and copying over the wreckage, and the app is now identified by its binary rather than by anything that mentions its path (#136, #139).

Terminal

• tmux windows are the panel's own tabs, addressed by window id rather than by tab position, so renaming or reordering cannot switch you to the wrong one (#154, #158).
• Tabs answer at once instead of taking seconds, and the panel now shows when tmux is waiting for the rest of a prefix key (#159).

Git

• Sync is no longer blocked on a branch that tracks the trunk directly. Its upstream IS its base, so "behind upstream" and "behind base" are the same commits and merging is what closes them: the old guard disabled the only working action and recommended a pull that could not run (#130).
• Handing a conflict to the chat now names the branch, the base, the worktree and which side is theirs, spelling out the rebase inversion that otherwise resolves an entire rebase backwards (#143).

Approvals and notifications

• A restart no longer auto-allows a hold a human was deciding on (#131).
• Approval requests survive a desktop Do Not Disturb setting agentglass does not own, and agentglass has its own quiet switch for mirrored notifications instead (#138, #152).
• The notch stops narrating old news, and holds jump the toast lane (#150).

Fleet and updates

• An open tool call now carries evidence of life: when the transcript last grew, or when the file a write tool named last changed. Reported only, because nothing classifies on it yet: the signal deserves watching against real runs before a state machine depends on it (#141).
• A newer release announces itself: a background check, a dot on the settings button while it stands, and one note per tag rather than a reminder every six hours (#145).
• Docker reads labels by name and samples only the containers on screen (#142).

Housekeeping

• Two test suites were reading the developer's real database; three source files carried NUL bytes that hid them from grep (#144, #146, #151, #155).
• package.json had said 0.2.0 since that release, so local builds misreported their own version (#161).

Thanks to @eatmonster, whose "check evidence of life, not liveness" framing is what the fleet work is built on.

v0.3.0

v0.3.0 — the workspace grows up: in-app conflict resolution, docker project controls, and a cockpit you can rearrange

Fifty-four commits since v0.2.1. The theme is the workspace becoming a place you operate from, not just watch.

Git is a real workflow now. Resolve merge conflicts without leaving the app, one conflict at a time. Undo the last merge while it's still cleanly reversible. Sync a checkout from its base, pull before sync, and sync is disabled while the tree is dirty so you don't paper over local work. The base picker offers the refs you can actually merge from, and file changes group by day.

Docker controls the whole project. Start, restart and stop an entire project's containers at once. A shell docked under the logs, a findable console, container rows that say what they're doing, and row actions that look like buttons.

The cockpit rearranges to fit you. Drag the rail to reorder views, one close per view, a single draggable sidebar width, rebindable view keys. A captioned notch over every view that mirrors desktop notifications, with a 12/24-hour clock and toasts. Panes flag what's only on this machine and fold away git-ignored files.

Settings, and an in-app updater. Tabbed panes and shortcuts that take any modifier combination. Update to a released tag from the app's own clone, no reinstall to move forward.

Desktop hardening. Serve the UI from a stable origin so preferences survive, an ephemeral static port so a second instance can start, setuid chrome-sandbox so the installed app launches, and the window shows before the server is up. Server-only installs got single-port deploys too.

Radar. The outer ring is anchored to the compaction threshold, so "edge" means "about to compact", not "model max".

Installers below: AppImage + .deb (Linux), .dmg (macOS arm64/x64), .exe (Windows). Still unsigned: right-click then Open on macOS, "More info" then "Run anyway" on Windows. Source install works as always.

Thanks

• @icey2488 — single-port deploys: the server can serve the built UI from the API port, so a headless box runs one process on one port (#115).
• @Solaris-star — the radar's compaction-anchored radius (#86) and two docs passes: the "Extending agentglass" guide (#87) and a README refresh (#88).
• @mgalore — made the tool-latency label say what it actually measures (#112).

MIT. Issues and PRs welcome.

v0.2.1

v0.2.1 — the desktop app moves to Electron (and stops stuttering)

The desktop app rendered the web UI in WebKitGTK (Tauri), which on Linux has no
GPU process and composites in software. With the window focused, the live radar
and streaming dashboard repainted continuously and burned ~115% of a CPU core;
the terminal visibly stuttered.

This release runs the exact same UI in Electron/Chromium, which composites on the
GPU. Same pixels, same features — the dashboard and terminal now render in real
time. Measured: a GPU process is present and idle CPU drops from ~115% (focused,
WebKit) to a fraction of that.

• Desktop shell migrated Tauri -> Electron; the Bun server still ships as a
  bundled sidecar and comes up with the app.
• Installers: AppImage + .deb (Linux), .dmg (macOS arm64/x64), .exe (Windows).
• The web UI itself is unchanged — this is a shell swap, not a redesign.

v0.2.0

Downloadable desktop builds, a real chat panel, and a project scope that actually holds.

Fifty-nine commits since v0.1.0. The headline is that you no longer need a Rust toolchain to run agentglass: this release ships installers.

• Linux — .deb
• macOS — .dmg, Apple Silicon and Intel
• Windows — .msi and .exe installer

The builds are unsigned. macOS will refuse the first launch: right-click the app and choose Open to get the override prompt. Windows SmartScreen will warn; "More info" then "Run anyway". If you'd rather not trust an unsigned binary, the source install still works exactly as before.

• Chat is a real panel now. Resume a finished session instead of only reading it, paste images into the composer, full-size layout, and a nudge when a chat is waiting on you. Long thinking turns no longer die to a 10s idle timeout.
• Sessions explain themselves. See what each tool actually answered, not just what it ran, and open a single subagent's thread on its own.
• An open project is a real boundary. Scope now constrains the dashboard, the docker panel and chat itself, rather than acting as a display filter.
• Project picker with filesystem path completion, and AGENTGLASS_PROJECTS_DIR accepts several roots.
• A settings modal in place of the "⋯" dropdown.
• New identity — a theme-reactive loupe-radar mark, a flight-ops landing page, and the live demo at /demo/.

Plus a long tail of fixes: stats totals, uptime accounting, diff theme loading, terminal SIGPIPE leakage, worktree session attribution, auth failure surfacing, and a mobile pass on both the dashboard and the landing page.

Thanks

Most of what got fixed this cycle came from outside the repo:

• @Solaris-star — made hook installation work on Windows (#23, #28), then wrote the test suite this project didn't have: pricing (#31), ingest (#32) and terminal command discovery (#33).
• @icey2488 — uptime counted from the server process rather than the oldest event (#69), ?window=all no longer zeroing every total (#70), and AGENTGLASS_PROJECTS_DIR accepting several roots (#71).
• @JGtm — reported the unquoted-path failure on Windows that started that thread (#15).

And @bergmorten, whose security work shaped v0.1.0 and still underpins the token auth and gate controls here.

MIT. Issues and PRs welcome.

v0.1.0 — the cockpit goes public

The first tagged release of agentglass: a real-time Mission Control dashboard and workspace for AI coding agents, across every provider and every project on your machine.

Highlights

• Machine-wide visibility. Every Claude Code session on the machine, live and historical (transcript scanner + hooks, deduped), plus an OpenTelemetry receiver for Codex CLI, Gemini CLI, LangChain and anything else that speaks GenAI telemetry.
• A workspace, not just a dashboard. Syntax-highlighted diff review of everything the fleet changed, a lazygit-style git panel, a docker panel, a real PTY terminal, and a chat panel that drives local Claude sessions.
• Project picker. Scope the whole cockpit to one project, or a folder of projects, from inside the app — the choice persists across launches.
• A radar that means something. Distance from centre = context window used. A blip drifting to the edge is about to compact.
• Answers over charts. Sessions show what they're actually doing ("running Bash · 8m"), long-running tools raise alerts, and the live feed has a lanes mode: one column per session.
• One-click project commands. Makefile targets and package scripts discovered across the whole repo (root and subfolders), grouped by folder, with descriptions.
• A security model that says what it is. Localhost-only by default, origin checks plus a DNS-rebinding Host guard, opt-in token auth for shared machines and network binds, an opt-in fail-closed gate, and a security test suite running in CI — huge thanks to @bergmorten for #1, this repo's first outside contribution.

Try it

• Live demo (fake data, no install): https://sirallap.github.io/agentglass/
• Quickstart: clone → bun install → bun run setup → bun run dev → http://localhost:6180
• Read the security model in the README before installing — it's short and honest.

MIT. Issues, ideas and PRs very welcome: the roadmap lives in the issue tracker.