MyOffice

What it does

MyOffice is a local control room for your AI agents. Tell the CEO what you need in plain words; it routes the job to the right helper and you watch the office while it runs. Each helper is an animal at a desk: it wakes when work arrives, shows its current action on its screen, and drops the result into Needs you. Idle agents sleep, so idle time costs zero tokens.

What it makes easier

• See what is running — every agent, its room, its current tool and your own terminal sessions on one screen.
• Stop scrolling terminals — each run keeps its prompt, duration and full result in History.
• Stay in control — nothing is published, mailed, paid for or pushed by itself; you approve each result.
• Give work in plain words — the CEO picks the right helper, so there is no prompt engineering.
• Review before shipping — a second helper returns APPROVE, CHANGES NEEDED or BLOCKED.
• Keep it private — the server binds 127.0.0.1; prompts, output and file contents stay on your machine.
• Pay nothing while idle — no work means no model calls and no tokens.

Highlights

A real office, not a dashboard	Isometric 3D rooms, animals at desks, clickable agents, a live event stream, a bottom activity dock and a daily token counter.
Any agent	Claude Code is automatic; OpenCode is read from its local database; Codex, Gemini and any other CLI plug in through myoffice run, a hook or a JSONL watcher.
Small and dependency-free	Python standard library + a single-page three.js frontend. No database server, no cloud account.
Built for a team	A CEO routes work to specialists (SOC Analyst, Lab Builder, Content Writer…) and a second helper reviews the result.
Made to be watched	Light and dark themes, sound with a mute toggle, focus mode, room and agent panels, and a click-anywhere overview.

A look inside

Click any image to open it full size.

Install

Try it without installing: open the live demo — fake agents, no backend and no model calls.

macOS or Linux (Windows via WSL2):

curl -fsSL https://raw.githubusercontent.com/halilbaris/AI-Office/main/install.sh | bash

This clones into ~/MyOffice, creates the private agents.json and bridge.json, and puts the myoffice launcher in ~/.local/bin.

myoffice start

Manual checkout

git clone https://github.com/halilbaris/AI-Office.git
cd AI-Office
./start.sh

Requirements: Python 3.10+, git, and one of Claude Code or OpenCode for the office's own agents. tmux is needed to give work from the office; watching your own sessions works without it.

Try the demo, no setup

cd frontend && python3 -m http.server 8000
# open http://127.0.0.1:8000/?demo=1

The demo uses frontend/demo-data.json: fake agents wake up, work and finish without a backend or any model calls. ?auto=0 turns automatic runs off and ?light=1 renders at lower resolution.

Bring any agent

Three ways in, all on the same local event stream:

1. Wrapper — works with any CLI

myoffice run -- codex exec "fix the failing test"
myoffice run -- gemini -p "summarize this repo"
myoffice run --agent writer -- ./my-agent

The wrapper tees the output to your terminal and sends only tool names and short targets to the office. It never changes the command's arguments or working directory.

2. Hook — for tools that can call a command

myoffice hook --agent writer --tool read --target README.md

3. Passive logs — for tools that already write JSONL

cp watch.json.example watch.json
# list the Codex, Gemini or other session folders you want to watch

OpenCode is detected automatically from ~/.local/share/opencode/opencode.db (v1 and v2 schemas). Claude Code is read through the built-in bridge (~/.claude/sessions + transcripts).

Every session carries its source (claude, opencode, wrapper, hook, log), so the UI can show where it came from.

How it works

flowchart LR
  B["Browser · 3D office"] <-->|JSON + SSE| S["server.py · 127.0.0.1"]
  S --> R["runner.py"] --> T["tmux"] --> C["Claude Code / OpenCode"]
  C -->|progress| S
  S --> BR["bridge.py · read-only"] --> CL["~/.claude"]
  S --> P["providers/ · read-only"] --> OC["OpenCode DB · JSONL logs"]
  X["myoffice run / hook"] -->|POST /api/ingest| S

• Frontend — a procedural three.js office (no photos or model files), a live event stream, the activity dock, room and agent panels, sound, light/dark and the demo mode.
• Backend — Python standard library, bound to 127.0.0.1. runner.py starts agents in tmux windows; bridge.py watches your own Claude Code sessions; providers/ watches OpenCode and configured JSONL logs.
• Events — events.jsonl is the single log. The browser replays it over SSE and the terminal list is merged from every source.

Privacy and safety

• The server is local: 127.0.0.1 only.
• The bridge and providers are read-only against agent tools.
• Only a tool name and a short target are stored. A shell command is reduced to its first word; prompts, model output and file contents are never written.
• Sessions are mapped to office agents through bridge.json (cwd → agent, default fallback).
• Nothing is published, emailed, paid for, or pushed by the office. The last step is always yours.

Configuration

File	Purpose
agents.json	Your private team: cwd, model, tools and preset tasks. Created from agents.example.json.
bridge.json	Maps a working directory to an office agent. Created on first run; edit it to match your folders.
watch.json	Optional passive JSONL watchers. Copy watch.json.example.
events.jsonl, state/, runs/, output/, memory/	Local working data (gitignored).

Agents, models and folders are also editable in Settings → Agents. Settings → Health explains anything missing in plain words.

Development

./scripts/check.sh

Runs Python syntax, the unit tests (59), JSON and frontend checks, the leftover-Turkish rule and an API smoke test on port 8798.

• Python standard library only; no new runtime dependencies.
• Add tests under tests/ for parsers, mapping and ingestion changes.
• See CONTRIBUTING.md.

Origin

MyOffice was originally taken from and is developed on top of selmakcby/ajan-ofisi. The original project provided the 3D office, the CEO flow and the agent runner. This fork adds the universal agent ingestion layer, the bottom activity dock, the installable CLI, the sound and overview controls, and this documentation.

License

PolyForm Noncommercial 1.0.0 — personal, noncommercial use. The original project's copyright and license remain with its authors; see the upstream repository.