Stop being the human in the inner loop. wta runs a fleet of AI coding agents in parallel — each in its own git worktree + tmux session — and gives you the harness to drive them: define a goal and a machine-checkable “done,” and wta re-prompts an agent until it passes. Switch between agents and hand off context from one dashboard, and see what each run costs. A single ~1 MB Rust binary that runs in any terminal.
Most tools stop at “spin up N agents in isolation.” wta is the harness around the loop — it drives agents to done, lets you move between them, and shows what they cost:
- Close the loop — give it a goal and a
verify.sh; wta re-prompts the agent until the tests pass, then locks the fix in as a check every future agent must clear. - Work across a fleet — run many agents at once, switch between them from one dashboard, and hand off one agent’s context into a fresh one.
- Analyze the run — per-agent tokens and estimated cost, with usage-over-time charts so you can see where the budget went and compare agents.
brew tap zakrad/wta && brew install wta # macOS / Linux
curl -fsSL https://raw.githubusercontent.com/zakrad/wta/main/install.sh | bash # prebuilt binary
cargo install --git https://github.com/zakrad/wta # from sourceNeeds tmux, git ≥ 2.20, and an agent CLI on your PATH (claude by default —
set WTA_AGENT_CMD to change). Add --features telegram for remote control.
cd your-repo
wta new fix-auth # new worktree + branch + starts the agent in a tmux session
wta # the dashboard — a live tree of EVERY repo's agentsBare wta opens a global dashboard across every repo you have agents in;
wta dash --here scopes to the current one. In it: j/k move · Enter attach ·
Tab Preview/Diff · i send a line without attaching · ? help. Try it free with
WTA_AGENT_CMD=bash wta new scratch.
Each agent runs in its own worktree (agent/<task> under .agents/) and its own tmux
session on a dedicated server (tmux -L wta), namespaced per repo, with a stable
WTA_INDEX/WTA_PORT_BASE so parallel dev servers don’t collide.
📖 Full per-feature manual → MANUAL.md.
Close the loop — the harness: a goal plus a machine-checkable “done,” and wta drives
the agent there. wta loop <task> re-prompts the agent with your .wta/verify.sh
output until it passes, with guards (--max, --no-progress, --timeout). wta lock <name> -- <cmd> freezes a past failure into a check every future agent must pass, and
wta cron add … --every <dur> fires the loop on a schedule — work while you sleep.
Work across a fleet — wta new (with --base, --model, --effort) and wta fanout <name> -n N (run N agents on one prompt, compare, keep the winner) scale you
out; the global dashboard and wta attach / i quick-send move you between agents;
wta handoff <from> <new> migrates one agent’s context into a fresh one, and wta send
/ wta board let agents coordinate.
Analyze the run — wta cost [<task>] --chart shows per-agent tokens and an
estimated spend with a usage-over-time chart (rate or cumulative, tokens or $) and a
model timeline, straight from the agent’s transcripts — no tracking overhead. wta supervise watches the fleet and escalates stuck / needs-input / crashed agents.
Review & merge — the Diff tab shows a colorized diff vs the agent’s base branch;
wta review <builder> spawns an independent maker/checker agent; wta push --pr opens
a PR against the agent’s base. (wta matrix also previews which branches conflict via
git merge-tree, read-only, if you want it before merging.)
Keys: n/N new · b new from a branch · s stop · D kill · p push/PR ·
v run checks · e open in your editor · J/K reorder · Shift+↑/↓ scroll ·
q quit. The Preview keeps the agent’s real colors — no need to attach.
Status glyphs: ⠋ running · ● ready · ▲ needs input · ◆ review (finished, unseen) · ✓ merged · ✗ exited.
wta shares the parallel-worktree substrate with tools like Claude Squad and Superset, but it’s built around driving the agent loop, moving between agents, and measuring the run — not just launching them.
| Capability | wta | Claude Squad | Superset |
|---|---|---|---|
| Loop an agent to green until the tests pass, unattended | ✅ | ❌ | ❌ |
| Hand off one agent’s context into another | ✅ | ❌ | ❌ |
| Per-agent tokens + estimated cost, usage charts | ✅ | ❌ | ❌ |
| Fan-out N agents on one prompt, compare, keep winner | ✅ | ❌ | ~ |
| Terminal-native — any terminal / over SSH, one binary | ✅ | ✅ | ❌ |
| Parallel agents in isolated git worktrees | ✅ | ✅ | ✅ |
| Visual side-by-side diff / open in any IDE | ❌ | ~ | ✅ |
Competitors move fast — verify current capabilities before relying on any ❌.
wta needs tmux and git ≥ 2.20 (native Windows means WSL). The core is
agent-agnostic — any CLI works via WTA_AGENT_CMD. Three conveniences are Claude
Code-specific and degrade gracefully otherwise: the ▲ needs-input status and
finish/needs-input notifications (Claude Code hooks), the auto-dismiss of the
folder-trust prompt, and the estimated $ in wta cost (token counts stay exact).
| Var | Default | |
|---|---|---|
WTA_AGENT_CMD |
claude |
agent CLI started in each session |
WTA_SKIP_PERMISSIONS |
1 |
run with --dangerously-skip-permissions; 0 or wta new --safe re-enables prompts — Claude only |
WTA_WORKTREE_DIR |
.agents |
worktree dir under the repo root (gitignore it) |
WTA_CONTEXT_FILES |
CLAUDE.local.md .env .env.local .mcp.json |
untracked files copied into each worktree (kept out of pushes) |
WTA_NOTIFY_SOUND / WTA_TMUX_NOTIFY |
1 |
finish/needs-input sound / in-terminal toast |
Full variable reference, per-repo .wta/ setup (setup.sh/verify.sh/teardown.sh),
isolation slots, and Telegram remote control are in MANUAL.md.
MIT