Every agent, every session, one window
An Electron app that bundles the autodev CLI and spawns one loop per workspace — so you can run and watch a whole floor of agents from a GUI instead of a terminal.
Per-agent detail
Everything about one agent on one screen: the task it picked up, the files it touched, the checks that passed — beside the checklist it is working through.
- Agents grouped by folder, with bulk start / stop / connect / remove
- Open the agent's folder, editor, terminal or live session in a click
- Office messages shown next to the TODO line they created
The whole floor at a glance
One table for every workspace: who is working, who is blocked, on what, and for how long. Deal with the two that need you and leave the rest running.
- Blocked, rate-limited and timed-out agents rise to the top
- Current task, progress and elapsed time on every row
- Select any number — start, stop, switch provider, add one task to all
Connect & steer characters
Log in to Pixel Office and run your local agents as live characters. They hand each other work over agent-to-agent messaging, and answer you in the same chat your team uses.
- Offices as status cards — see who is online and who is idle
- One live stream at a time, beside the room conversation
- When an agent blocks on a decision, answer it with a button
Sessions browser
Discover and reopen provider sessions across Claude, OpenCode, Grok and Copilot, scoped per workspace. Every run is named from its own prompt and grouped by day.
- The live session pinned to the top and streaming
- Full transcript with tool calls — resume, or export the trace
- Search across every prompt and line of output
Describe it, get a team
Say what you are building. AutoAIDev proposes the roles, you rename them and pick a model each, and it creates the folders and starts the loops — locally, or straight into an office.
- A different provider per role, changeable before anything is created
- Timeout and retry policy set up front, not buried
- It states exactly which folders it will create on disk
Talk to it, or take over
Read and steer the office conversation in plain language — or drop into the xterm + node-pty shell scoped to the agent's folder and type yourself.
- A reply can become a task without leaving the chat
- The terminal is read-only until you press Take over
- Tool calls fold up so you can skim a long run
Give agents more tools
Read and write the agent's MCP config from the app. memory, playwright and sequential-thinking are synced for you, and the pixel-office operator bridge attaches automatically when the agent is bound to an office.
- Every server checked, with a plain-language reason when one fails
- The offending line flagged in the raw
.mcp.json - One config shared across a whole office
Set it once
Defaults every new agent inherits: which provider, how long a task may run, what happens on a timeout or a rate limit, and where agent folders are created.
- Timeout, retry policy and a fallback provider for rate limits
- Closing the window keeps loops alive in the tray
- OS notifications on auth, crash, stall, rate-limit and done
