Online · working 3 issues · 0 humans awake

The AI dev team that ships while you sleep.

Open a task. Pick a workflow, an agent, and a model — then hit ▶ Play. A crew of AI agents plans it, splits big work into sub-issues, builds, tests, reviews its own diff, and self-heals CI. Point and click, no tags. Then go play padel.

Put the crew to work Watch a ticket flow ↓
scroll
Runs on any model, any provider — your keys, your box
// the crew

Specialists, one click away.

Each role is a real agent with its own model, tools, persona, and skills. Pick one from the roster, pick a model, hit Play.

Every role is fully editable — persona, model, tools, skills, avatar. See the Agent editor ↓

// the layer above the crew

An orchestrator that runs it. A watchdog that checks it.

Above the per-task agents sit two always-on minds — one drives the whole agency across every repo, the other keeps it honest from the outside.

The Orchestrator

agency-wide + per-repo · always aware

Spans your whole agency and every repo it watches — it always knows what's running, what's blocked, and what's next. It cross-checks runs, plans with you in dialog, or takes the wheel.

🗣 Dialog plans and steers with you, one message at a time.
🤖 Autopilot auto-resume after limits, auto-merge clean PRs — the board keeps moving while you're out.

The Analyzer

outside the agency · read-only

A separate process that watches from the outside and keeps the crew in check. Read-only on the telemetry, it can't touch the agency — it only opens advisory issues proposing new skills, hooks, and deterministic code.

🔎 Watch reads failures, cost, loops, and tool usage across runs.
📋 Advise files improvement issues — you decide what ships.
// one task, start to ship

Pick a workflow. Hit Play. Watch it ship.

Point and click — no slash commands, no tags. Choose a workflow (or a single agent) and a model, press ▶ Play, and watch the run stream live. Or close the laptop and go play padel.

▶ Play ✓ Approve ⤴ Merge ⏸ Interrupt
// big work, decomposed

One chat becomes an epic — and a stack of sub-issues.

Talk through something large. The Decomposer breaks it into linked sub-issues, each run by a full agency pass. A final integration review checks the merged whole, then one Merge click ships every child PR.

💬 chat / issue
Build a billing dashboard: schema, API, and UI.
Decomposer → that's an epic. Proposing 3 sub-issues…
🧩 epic #128 · checklist
✓#128a data schema merged
▶#128b REST API building
•#128c dashboard UI queued
🚀 integrate & ship
⌾ integration review on merged main
⤴ Merge ships all child PRs at once
route: {files: db/*} each sub-issue is its own full agency run — concurrency ×3 disjoint file footprints run at the same time.

The built-in ▶ Play train auto-starts each child when the previous one merges; parallelism kicks in when footprints don't overlap (file-lock gated).

// make it yours

Design the pipeline. Rewrite the agents.

A visual workflow builder and a full agent editor — no code required. Wire custom flows, gate them, and give each role its own brain.

Workflow builder · release-pipeline

Wire a custom pipeline

Chain agent step-nodes with SVG connectors and gate badges — each step gets its own model, forced skills, hooks, and budget.

Planplanner · opus-4.8skill: research
✓ continue
Builddeveloper · sonnet$4 budget
tests-fail?
Testtester · haikuhook: lint
↻ loop → Build (max 2)
Reviewreviewer · sonnetgate: approve
changes-requested → stop & ask you
Agent editor · Developer

The Developer

agent: developer · role: builder
You build on a fresh branch with atomic commits. Reuse before you create. Follow the house playbooks to the letter. Terse output; plain English only for the human.
sonnet  ·  fallback → glm-4.6
Can write code Read · Write · Edit · Bash · Grep
atomic-designtddreuse-firstkiss
// token economy

Agents talk in shorthand. To you, plain English.

A strict output style strips preamble, summaries, and narration between agents — so cost goes to the work, not the chatter. Only the human-facing reply is written out in full.

❌ typical agent~180 tokens
Sure! I'll now go ahead and analyze the authentication middleware. First, let me carefully examine the token expiry logic to understand what might be causing the issue you're describing. After reviewing, I believe the problem is likely related to how the comparison operator is used…
✓ dev agency~24 tokens
auth mw bug. expiry check uses `<` not `<=`. fix L42. re-test → green. handoff → reviewer.

Backed by per-role playbook loading (only what a role needs), capped lesson injection, and skipping redundant re-reads. No secret codec — just discipline that compounds across thousands of runs.

// the machinery

The boring-but-brilliant parts.

What turns “cool demo” into “runs our backlog unattended.”

// bring your own brains

Two runtimes. Every major provider.

Cheapest model that can do the job, per role — Planner on Opus, builders on Sonnet, testers on Haiku. Swap any of it for GLM, DeepSeek, Gemini, Grok, and more. Models are discovered live, so if a provider lists it, you can route to it.

runtime

Claude Agent SDK

Anthropic's official agent runtime — Claude subscription token or API key.

runtime

pi-coding-agent

Runs in-process and unlocks every provider below, plus OpenRouter — any model they expose.

// always something happening

A peek at the agency floor.

A simulated feed. The real dashboard streams every agent's steps, checks, and commits to your phone — and mirrors the whole conversation to GitHub.

agency://live-activity
// rave reviews†

Teams that stopped babysitting the backlog.

†Dramatised mock-ups for this promo. Any resemblance to your exhausted on-call engineer is coincidental.

// it's yours

Self-host it in minutes. Zero env vars.

Deploy on your own Coolify box — a master key auto-generates, everything else is an in-app wizard. These “tiers” are a playful mock; it's one app that scales with the models you plug in.

Your backlog won't ship itself.
Or… will it?

Open an issue tonight. Wake up to a draft PR. Say “ship it” before coffee.

Get it on GitHub ★ Meet the crew again