Performance

A desktop app that stays light

TAO CODING next to Orca (Electron desktop app) and Herdr (terminal UI). Same machine, plain shell panes, lower is better on every chart.

TAO CODINGOrca · HerdrPtyxis, the terminal Herdr runs in

How we measured

Machine: AMD Ryzen 5 5600H, 18 GB RAM, Ubuntu 26.04, GNOME on Wayland
Versions: TAO CODING 1.0.4 built from source without the account sign-in (a clean profile cannot sign in) · Orca 1.4.217 from its .deb · Herdr 0.9.3 release binary inside Ptyxis, the default Ubuntu terminal
Setup: Each app gets a fresh profile and its own session bus, restores a saved layout of plain bash panes, and keeps its default window size. Background daemons count as part of the app.
Open: From launch until the app has a shell and its processes go quiet (under 25% of one core for 2 s), so window, renderer and terminal are all done. Median of 10 after one warm-up.
Memory: Sum of PSS over the whole process tree after 60 s idle, with 1 and 6 bash panes, median of 3.
Output: cat of a 50 MB coloured log file in one pane, timed until the app goes quiet again. Timing only the cat would measure how fast the backend reads the pty, not how fast the terminal draws. Median of 5.

Orca runs on the desktop session bus: on an isolated bus it waits about 25 s for a keyring at every launch, which is our setup, not Orca. Most numbers are the median of 3 to 10 runs; Orca's open time is 5 runs, Herdr's memory 1 run at each pane count and its output 3 runs. We publish every number, including the ones we lose. Linux only.

One agent runs fine anywhere. Five agents break an ordinary terminal: tabs hide who is doing what, agents overwrite each other, and you still check the page and the phone yourself.

Run a team

Split panes with live state, one git worktree per agent, and a room where agents talk to each other while you moderate.

See what users see

The web preview and your Android phone sit next to the terminals. Point at what is wrong and hand it to an agent.

Agents use your tools

Through the built-in MCP server, agents open previews, read network logs, query the database and spawn worktrees.

Your agent team

Composer

Start a team in one step

Pick a folder, add agent rows, turn on worktrees and create. Panes open with live state: working, blocked, done.

Claude Code, Codex, Gemini and 8 more
Up to 12 panes per workspace

Worktrees

Every agent on its own branch

Each agent pane starts inside its own git worktree, so parallel edits never collide. Review diffs and check merges in the Git tab.

Inline diff, ahead and behind
Dry-run merge check

Meeting Room

Agents talk to each other. You moderate.

Drag panes into a room. Agents probe each other or hand over a prompt to run. You message one agent or the whole room.

Messages wait while you type, never dropped
Drag a pane out and its access ends

Built for web and mobile

Website

Point at the UI, the agent fixes it

Pick an element in the live preview and drag it into a terminal. The agent gets the selector, styles and source file and line.

Phone and tablet viewports
Network log and console the agent can read

Mobile

Show the agent what the phone shows

Mirror an Android phone or emulator. Cut a region of the screen and drag it into a terminal as context.

Tap, type, Back, Home
Works with emulators

Database

Your data, next to your code

A table client for Postgres, MySQL and SQLite. Browse, add and edit rows, and choose what each agent may do with them.

Full, read-only or no agent access
Every agent query is logged

Architecture

See the system your agents are changing

An interactive diagram of the project. Click a box to see its files. Boxes light up in the agent's colour while it edits them.

Stale areas are marked
New projects get a design before any code

Right panel

Also in the box

Attention

Know who needs you

Blocked agents turn red. One shortcut jumps to the one that has waited longest.

Broadcast

Type once, reach every agent

Send one line to every pane you pick: run the tests, commit, stop.

Restore

Close the app, keep the team

Workspaces, agents and worktrees come back, and conversations resume.

Network

Request log for you and agents

Headers, payloads and responses under the preview. Credentials are redacted.

MCP

No setup for agents

Every pane connects to the built-in MCP server. Claude Code, Codex and Qwen are enrolled automatically.

Status

Context and quota

Context use per pane and 5-hour and 7-day quota chips, read from local files only.

FAQ

Is it free?

Yes. You sign in with a free account (Google, GitHub or email). You bring your own agent subscriptions or API keys.

What does TAO CODING send home?

An anonymous app-open event and a 5-minute heartbeat with the app version and OS. Nothing from your code or terminals.

Do I need to set up MCP?

No. Every pane gets the server address in its environment. Claude Code, Codex and Qwen are enrolled automatically.

Is there an iPhone app or iOS mirroring?

No. The Mobile tab mirrors Android phones and emulators only.

Are there performance numbers for macOS and Windows?

Not yet. The published numbers are from Linux. The benchmark scripts are public if you want to run them.

The agent workspace for web and mobile developers.