Built and working
- Agent intake, previews, visual diff and verdicts
- Attention inbox and review links
- Promotion and grafts with backup tags
- An optional built-in agent for Rails apps
Software teams
BranchGlass
Live previews, visual diffs and one-key verdicts for what AI coding agents build, with one view of what needs a person across every app.
See it work
Real screens from the working build, judging sample apps.
How it works
- 01ReportHooks for Claude Code, Codex and Antigravity report each finished turn, without ever slowing the agent down.
- 02PreviewEvery version runs in its own isolated preview, with its own copy of the database.
- 03CompareSide by side, live, or as a pixel diff that numbers each changed region.
- 04DecidePick or reject with one key, then promote the winner to a branch with a backup tag.
Features
01
One view of what needs you
Agents waiting on a question, reviewer feedback, broken previews and versions ready to judge, across every project, sorted by four fixed rules.
- Works across all your projects at once
- Snoozed items return when something changes
- Named, time-limited review links for teammates
02
See exactly what changed
A pixel comparison shows how much of the page changed and numbers each region, with a heatmap and zoom. Side by side, overlay, wipe, blink and live views are there too.
- Live apps side by side, with synced routes and scrolling
- Anti-aliasing noise ignored
- Each changed region one click away
03
Know where each version came from
Each version records the agent, model, session and commit that produced it, and the agent's last message, beside a screenshot of the result.
- The same code reported twice counts once
- Provenance hidden from outside reviewers
- Grounded in the exact git tree
04
Promote without surprises
The winner moves to a branch fast-forward only, with the old tip tagged as a backup first. A branch that is checked out or mid-rebase is never moved.
- Compare-and-swap branch updates
- Grafts without touching your working tree
- A conflict writes nothing
Details
- Agents
- Claude Code, Codex and Antigravity hooks, a git hook, a command-line tool and an MCP server.
- Previews
- Each version in its own worktree and process on its own local address, within a memory budget. Idle previews stop.
- Databases
- A SQLite copy for every preview, made per schema. PostgreSQL cloning is tested with stand-ins.
- Stacks
- Rails is detected automatically; other apps run with a custom start command.
- Platforms
- macOS. Linux is covered by automated tests; Windows through WSL2 is not yet validated.
- Security
- The studio listens only on your machine. Review links are named, time-limited and revocable, with an optional macOS sandbox for previews.
Where it stands
A working build with 389 automated tests. macOS first.
Next
- Tests run on agent output
- A grid of two to four versions
- Review pins that become the agent's next task
- Blind comparisons and rankings
Common questions
Does it run my agents?
Your agents keep running where they do; BranchGlass is where you judge what they built. It also has an optional built-in agent for Rails apps.
Does it capture my prompts?
No. Agent hooks report when a turn ends. The intent behind a version is recorded only when you or the agent send it through the command line or MCP.
Which stacks does it support?
Rails is detected automatically. Other web apps run with a start command you set; dedicated adapters for other stacks are planned.
Does my code leave my machine?
The studio runs on your machine and listens only locally. Review links share snapshots with named reviewers for a set time.
See it with your own workflow
Tell us which agents and apps your team uses, and we can show BranchGlass judging their work.