Run a team of AI coding agents from one terminal.
wdx keeps Claude Code, Codex, Cursor and Antigravity agents running in their own git worktrees, long after your terminal closes. TIDE, its Terminal IDE, shows all of them in one window, with their diffs, plans and code.
TIDE is a wide-screen tool: swipe the window sideways to see all of it.
A replica of TIDE, scripted in your browser: no AI CLI, daemon or git runs here.
Every agent, worktree and plan of a project, one key away.
These are TIDE's own screens, drawn live by the same engine as the window above. Open any of them to try that screen.
The TIDE pages, screen by screen →Built for a day of running agents in parallel.
Know at a glance who needs you.
wdx reads each agent's screen, so TIDE can mark every row working, waiting or idle, whichever AI CLI it runs. A question from an agent turns its row amber until you answer.
- Agents run in the daemon: close TIDE, or the laptop lid, and they keep going
- Child agents, restarts that resume the same session, and the context size on every row
- Answer, steer or send to an agent from TIDE, the CLI or a script
A branch per agent, and a diff before anything lands.
Give an agent a worktree and it works on a branch of its own. TIDE shows what each one has committed and what it has not, file by file, then lands it on main after your landing check, or opens a PR.
- Rebase onto main when asked; conflicts go to an agent you pick
- git is read from events, never on a timer, and never writes your index
Split a feature into tasks, and let wdx run them.
A plan is a DAG of tasks. wdx starts an agent for each task that is ready, each in a worktree of the plan's, and lands every finished task on the plan's branch. Plans of plans land as stacked PRs.
- Drawn the way
git log --graphdraws commits - A task that fails waits for you to retry, land or skip it
Docs with revisions
Designs live in wdx with a status, wip to approved, so people and agents read the same thing.
The docs guide →For scripts and AIs
Every command returns one JSON object with stable error codes, and never prompts.
The scripting guide →Terminals that stay
Shells in PTYs that outlive the daemon, with their scrollback, attachable from anywhere.
The terminals guide →Observable
Every run logged with a debug id, metrics in Grafana with one command.
The observability guide →TIDE for people, the same features for scripts and agents.
Agents that wdx runs learn it from wdx help agents: they make child agents, worktrees, docs and plans through wdx, so you see all of it in tide.
$ wdx agent create --name fix-auth --worktree fix-auth --aca codex created agent fix-auth in shop-api worktree fix-auth (new) $ wdx agent list --json {"agents": [{ "name": "fix-auth", "aca": "codex", "state": "running", "activity": "waiting", "prompt": {"kind": "yes_no", "question": "Do you want to proceed?"}, "context": {"tokens": 154210} }]} $ wdx worktree land --name fix-auth landed worktree fix-auth on main
One daemon holds everything. Every window is a client.
TIDE and the CLI talk to the daemon over a Unix socket. Each terminal and agent lives in a holder process of its own, which keeps it running when the daemon restarts or is updated.
How it works, in depth →Your machines, one account.
Log your machines in to a hub you run, and reach each one's agents, terminals and TIDE from any other with --node. The hub passes the connection on but cannot read it: TLS runs between the two machines, and the machine checks its own authenticator code.
- Remote access is off on a machine until you set up its code there.
- A project's docs can be pushed to the hub, encrypted at rest.
- Self-hosted, for you or your organisation: a Linux server with one command, or Fly.io.
$ wdx hub login --url https://hub.example.com logged in to https://hub.example.com as alice; this machine is node desk $ wdx hub node list NAME ONLINE LAST SEEN VERSION OS ID desk (this) yes 2026-11-08 15:33:10 0.5.0 linux/amd64 node-1a2b3c4d laptop yes 2026-11-08 15:32:58 0.5.0 darwin/arm64 node-5e6f7a8b $ wdx tide --node laptop on node laptop code for laptop: 492133
The hub is optional: wdx works the same without one. There is no hosted hub to sign up to; you run wdx-hub yourself.
How wdx is built.
Four ways running several AI agents goes wrong, and how wdx fixes each
Lost terminal tabs, colliding git checkouts, silently blocked agents waiting on prompts, and mystery diffs. Here is how wdx fixes the four classic multi-agent failure modes.
2026-11-12 · acas, claude-code, codex, cursor, antigravity, multi-agentClaude Code, Codex, Cursor and Antigravity under one tool
Every AI coding CLI comes with its own terminal quirks, session formats, and permission models. Here is how wdx normalizes Claude Code, Codex, Cursor, and Antigravity into one uniform agent fleet.
2026-11-12 · story, productivity, agentsFrom one agent to five: what changed in my day
Before wdx: one AI CLI and a row of terminal tabs I lost track of. After: named agents, children under parents, several AI CLIs at once, all on one screen.
Four commands to your first agent.
The installer puts wdx in ~/.local/bin, adds shell completions and runs the daemon as a systemd or launchd user service. wdx needs git; each agent needs its AI CLI installed.
curl -fsSL https://raw.githubusercontent.com/srjn45/wdx/main/scripts/install.sh | shwdx project open --local ~/code/apiwdx agent create --name fix --worktree fixwdx tide