Twenty stages.
Each with its own layout, light, silhouette and sound. Agent Theatre reads what your repo is and suggests the ones that suit it.
The repo is the stage.
Folders are districts. Files are buildings. Test files are beacons. The agent walks to what it reads, builds on what it edits, runs the tests at the beacons — and chases gremlins when they fail.
12
A cast.
The lead is chosen per repo. Subagents arrive at the dock as other members of the cast.
0
Linear tweens.
Every motion is eased. A test in the repo makes sure none sneak in.
0 KB
Audio files.
Every sound is synthesized at runtime. Every stage has its own voice and ambience.
One key light.
Long soft shadows, ambient occlusion, rim light, aerial haze, cloud shadows drifting over the ground. Bloom only where something glows.
Six episodes. Six kinds of work.
Research, a new feature, a bug fix, a refactor, a deploy, a team of helpers — each a different repo on a different stage. Every one is synthetic: hand-written, never a captured session.
Bug fix
The login timeout
A 15-minute session drop. Two failing tests, one grace window, a green run — and the evidence chip to prove it.
Research
The research brief
Four field notes, a data file, two web searches, one written brief. No tests to pass, so none claimed.
New feature
The export endpoint
Two new buildings rise, a route is wired, the suite is green first time.
Refactor
The rename
Six files touched, one test catches a slip, fifty-eight pass.
Deploy
Ship 1.4.0
All shell work at the forge: build, plan, a timeout, a retry, a rollout, a tag.
Helpers
The audit
Two helpers arrive at the dock and fan out while the lead keeps working.
The honest detail
If the agent says it's done, and no test has passed since the last edit —
a gremlin hides behind it.
No other visualiser checks the claim against the evidence. Agent Theatre shows the passing test count as a chip when it exists, and the gremlin when it doesn't. Nothing is invented; every beat traces back to an event.
Twenty stages.
One repo, twenty ways.
The same project is a ring island in one set and a terraced spine in another — and always the same in the same set. Scroll, then open any of them.
IslandA warm little island at golden hour.
OrbitA quiet station turning above a blue planet.
MetropolisA warm night city, lights just coming on.
HarbourFog lifting off a working port.
ObservatoryA mountaintop under a hard, clear sky.
GreenhouseGlass, leaves and warm afternoon light.
WorkshopA village of makers, sawdust in the air.
ArcadeNeon rain on a late-night strip.
KingdomStone walls, banners, a river below.
OutpostA desert station at dusk, dust in the light.
SnowfieldBlue hour, fresh snow, chimney smoke.
CanopyTreehouses strung through an old forest.
LagoonTurquoise shallows and white sand at noon.
VolcanoBlack rock, red sky, light from below.
CloudtopA floating town above a sea of cloud.
ReefUnderwater, where the light comes in ribbons.
BazaarOchre walls, striped awnings, late light.
MonasteryMist on the mountain, a bell somewhere.
MoonbaseGrey dust, black sky, one blue marble.
LanternsA night festival, paper light everywhere.
A cast of twelve.
Every character is drawn by one procedural routine from a short spec — a shape, a topper, a pair of eyes, a palette. No sprite sheets. Their light shows their mood.
How it works.
- 01
Events
Claude Code hooks (live) or the session transcript (replay), normalised into one small schema:
read_file,edit_file +n −m,run_command,test_result,subagent_start… Counts and paths. Never contents. - 02
Stage
The file tree becomes a deterministic arrangement — six layouts, chosen from the repo and the stage. Every file the session touches is always placed.
- 03
Director
Events become beats — walk, read, build, run tests, spawn gremlin, chase, celebrate — with a mood: curious, focused, lost, happy.
- 04
Renderer
Isometric WebGL. The camera follows the character; the whole stage appears only to open, celebrate and close, so it never gets cluttered.
prefers-reduced-motionrespected.
Privacy. Not a setting.
Agent Theatre runs entirely on your machine and talks only to 127.0.0.1. Events carry repo-relative paths, line counts, a command's first word and a classified intent — never file contents, prompts, command bodies or search queries. A test plants marker strings in all of those and asserts zero leaks. Hooks install non-destructively with a backup, and uninstall byte-for-byte.
Two commands.
$ npx agent-theatre hooks install
$ npx agent-theatreOpens the stage for the repo you're in. Run Claude Code in another terminal and watch. npx agent-theatre replay plays back any past session. Node 22+.
VS Code
The same stage inside the editor, from the same hooks. Coming to the Marketplace.
Open source
MIT. Core, CLI, renderer and this site in one repo. github.com/shreyassaw5394/agent-theatre
Part of a suite
LastSearch for research with evidence. Debug Anything for runtime observation. Agent Theatre for watching the work.