⚡ Early access / beta — shipping updates daily. Things move fast. Jump in now and grow with it.
Most AI coding tools give you a smart assistant.
You ask. It answers. You ask again. It answers again.
There's no process. No structure. No handoffs. Just you, manually steering an LLM through every decision from idea to shipped feature — one prompt at a time.
What if you didn't have to steer at all?
🎬 Watch the pipeline run - Coming soon..
📸 [GIF: intent typed into @orchestrator → routes to @prd → PRD writes itself → pipeline-state.json updates in real-time → next agent auto-opens]
That's not a demo environment. That's just VS Code + GitHub Copilot + junai installed.
The idea: a real engineering process, run by AI
junai is an agentic pipeline inside VS Code. Every stage of your delivery lifecycle has a dedicated specialist agent. The Orchestrator routes between them. State is tracked in a plain-text file in your repo.
Intent → PRD → Architecture → Plan → Implement → Test → Review → Shipped ✅
↑ every stage has a specialist, every transition is logged, every gate is explicit
You describe what you want to build. The pipeline figures out what needs to happen and who handles each part — automatically.
🤖 Autopilot: sign off once, watch it go
📸 [GIF: autopilot mode — pipeline-state.json updating on disk, next agent auto-opening in Copilot chat, handoff prompt firing, no mouse clicks]
Switch to autopilot mode and junai's watcher monitors pipeline-state.json in real-time. The moment a stage completes:
- Reads the routing decision
- Automatically opens the correct specialist agent
- Sends the handoff prompt — the next agent starts immediately
No copy-paste. No "now go tell the architect agent...". It just goes — until it hits a decision that genuinely needs you.
You sign off the intent once. The pipeline runs the rest.
24 specialists. Right model for the right job.
| Agent | Model | Why |
|---|---|---|
| Orchestrator, Architect, Anchor | Claude Opus | Highest-stakes reasoning |
| Code Reviewer, Debug | Claude Sonnet | Judgment over generation |
| Implement, Tester, SQL, Frontend | GPT-5.3-Codex | Pure code output |
| Diagrams, UI/UX | Gemini Pro | Visual + design |
9 MCP tools — pipeline control from chat
"Where are we?" →
get_pipeline_status
"Skip this stage" →skip_stage
"Who handles testing?" →get_agent_for_stage
3 modes for 3 situations
| Mode | What it does |
|---|---|
| 🎛️ supervised | You approve every gate — full control |
| 🤝 assisted | AI recommends, you decide |
| 🤖 autopilot | Sign off the intent once. Pipeline runs itself. |
Try it in 60 seconds
- Open VS Code (GitHub Copilot required — the one you already pay for)
- Search junai in Extensions, or → install from marketplace
- Open Copilot Chat → type
@Orchestrator→ describe what you want to build
The agent pool lives in .github/ — portable across machines, shareable across teams.
⚠️ This is moving fast
junai is in active beta. Features ship daily. The core pipeline is late beta and battle-tested, but rough edges exist.
If something breaks — open an issue. Fixes land same day or next. Early adopters shape what this becomes.
Top comments (0)