SELF-HOSTED · OPEN SOURCE
FLOCKBOTS.
A flock of specialized AI agents that ships production code for you — from a chat message on your phone to a QA-verified deploy.
PM · UX · DEV · REVIEWER · QA // ORCHESTRATED, GATED, AUTOMATED
INSTALL · 30 SECONDS · MACOS / LINUX / WSL
~10 MIN TO A RUNNING FLOCK
$
curl -fsSL https://raw.githubusercontent.com/pushan-hinduja/flockbots/main/install.sh | bash
$
flockbots init
$
// 01 · WHAT IT IS
01 / 04
Not one AI writing code. A team of specialized agents — each with a defined role, system prompt, and tool set — handing work off through disk-persistent artifacts. A coordinator picks the right agent at the right stage and enforces gates between them. Self-hosted. Authoritative state on disk. Your code never leaves your machine.
// 02 · THE FLOCK
02 / 04
A1
PM
Researches once. Writes a context pack every other agent reads. Picks model + effort per task.
A2
UX
Reads your design skills. Picks existing components, writes layout + interaction notes.
A3
DEV
Isolated git worktree. Tests, lints, typechecks. Opens a PR as a separate GitHub App.
A4
REVIEWER
Fresh Claude session, second GitHub App. Approves or requests changes — visible in the PR UI.
A5
QA
Drives staging in headless Chromium. Screenshots pass/fail and pings you via your communication channel.
// 03 · WHY A FLOCK
03 / 04
VIBECODING
One AI · one chat
- Re-reads your codebase every session
- Self-reviews its own code — bugs slip through
- Says "done" — you still have to manually QA
- Crashes lose intermediate work
- Burns tokens on broad greps in big repos
- Locked to your laptop — no shipping from a phone
FLOCKBOTS
A coordinated team
- PM writes one context pack — paid once per task
- Reviewer is a separate identity with a fresh session
- QA drives the deployed staging URL with Playwright
- SQLite-backed queue — crash-safe, resumable
- Knowledge graph — ~5–10× cheaper symbol lookups
- Telegram, Slack, WhatsApp — your phone is the UI
// 04 · THE PIPELINE
04 / 04
◌
YOU
PM
CONTEXT
UX
DESIGN
DV
BUILD
RV
REVIEW
PR
MERGE
QA
VERIFY
✓
SHIPPED
// FEATURES · WHAT'S IN THE BOX
+++
F.01
Knowledge graph
Indexes your repo into a symbol + import + call-site graph via graphify. Agents query the graph instead of greping — ~5–10× cheaper lookups on real codebases.
F.02
Rate-limit aware
Parses Claude retry timestamps, pauses sessions until they clear, defers L/XL tasks during peak hours. Queue 50 tasks, walk away — the flock paces itself.
F.03
Live office view
Pixel-art office where each agent sits at their desk. Watch them think, write, review and hand off in real time. The whole flock in one glance.
F.04
Phone-first chat
Telegram, Slack, WhatsApp out of the box. Natural language router (Claude Haiku) parses intent — no slash-command memorization. "merge the latest PR" just works.
F.05
Use your Claude sub
OAuths through the Claude CLI exactly like Claude Code. If you have Max/Pro, the entire flock runs on it — no API keys, no per-token bill. Or use an API key if you'd rather.
F.06
QA with screenshots
Playwright drives the deployed staging URL. Screenshots both states, uploads to Supabase Storage, and texts you the result. On failure, auto-creates a fix task.
A FLOCK OF AI AGENTS. SHIPPING CODE.
CHAT IT IN · WAKE UP TO A QA-VERIFIED PR