modu-ai/moai-adk

MoAI‑ADK – a verification‑driven harness for AI‑code agents

What it is

• An open‑source command‑line tool (written in Go) that orchestrates Claude Code (and other LLM back‑ends like GLM and Codex) as reliable software‑development agents.
• It adds a strict verification layer: every claim an agent makes (e.g., “tests passed”) must be backed by the actual command that was run and its output. Unverified claims are rejected at the system level.

Why it matters

• Large‑language‑model agents can generate code, but without external checks you cannot trust the results. MoAI‑ADK enforces a SPEC lifecycle (plan → run → sync) and a set of “TRUST 5” gates that bind evidence to each step.
• It lets you run many agents in parallel ("lanes") while keeping each lane isolated in its own Git worktree, preventing accidental overwrites.
• The tool is built for long‑running sessions: progress is stored in progress.md, and sessions can be resumed after a /clear without losing context.

Key concepts

Term	Meaning
Leader	A coordinating session that watches the backlog queue and hands out cards (units of work) to lanes. It never writes code itself.
Lane	A worker session + isolated worktree that processes a single card through the full SPEC pipeline. Multiple lanes run side‑by‑side.
Card	One unit of work (e.g., a bug fix, feature, or design change). Added via moai gtd "…".
Factory Mode	The default orchestration mode where a leader and a set of lanes collaborate.
SPEC	The three‑stage lifecycle: plan (author a specification), run (implement with TDD/DDD), sync (create docs/PR).

How it works

1. Add a card – moai gtd "fix stale rename hint" puts a task in the backlog queue.
2. Start the factory – moai cc -f launches the leader (the “factory”).
3. Spawn lanes – In separate terminals run moai cc -l (Claude), moai glm -l (GLM), or moai codex -l (Codex). Each lane receives a card, creates an isolated Git worktree, and runs the SPEC stages.
4. Verification – When a lane finishes a stage, it writes evidence (command output, test results) to progress.md. The leader only advances the card after reading this evidence, never on the lane’s claim alone.
5. Completion – After the sync stage the lane asks the user to run /clear; the user‑typed command confirms the work is truly finished.

Main features

• Factory Mode with automatic lane numbering and automatic run‑record handling.
• Parallel‑safe isolation: each lane gets its own worktree and branch (WT‑<slug>), and only one write‑capable agent runs per repository at a time.
• Cross‑session messaging control via crosssession.yaml (approve, reject, or auto‑accept messages from other sessions).
• Home‑directory hygiene: moai clean --home removes old state under ~/.moai; the location can be overridden with MOAI_HOME.
• Web console (moai web) that visualises lanes, SPEC counts, and queue status.
• Support for 16 programming languages via marker‑based auto‑detection, with equal tooling for each.
• Self‑improvement loop: failure patterns are recorded in a decision‑memory and can be promoted to new rules after user approval.
• Multi‑backend: choose Claude, GLM, or Codex per lane; the leader can run on a cheap backend (GLM) while lanes use more expensive models.

Getting started (quick)

# Start the leader (factory)
moai cc -f

# In separate terminals, start lanes (they will be auto‑numbered)
moai cc -l      # Claude lane
moai glm -l    # GLM lane (optional)
moai codex -l  # Codex lane (optional)

Add work with moai gtd "your task" and watch the web UI (moai web) or use CLI commands like moai factory status to monitor progress.

Who should use it

• Developers who want to experiment with LLM‑driven code generation but need strong safety guarantees.
• Teams that run Claude Code (or similar agents) at scale and want reproducible, auditable workflows.
• Anyone interested in building autonomous coding loops that respect token budgets, avoid runaway sessions, and keep a clean Git history.

Where to learn more

• Official docs: https://adk.mo.ai.kr
• Detailed “Factory Mode” guide (linked in the README).
• Community Discord: https://discord.gg/Z7E7Mdc5aN

---

All information above is taken directly from the repository’s README; no additional features have been inferred.