JCLAW is a high-fidelity orchestration layer for Google Jules, designed to transform the raw "muscles" of automated coding into a robust, context-aware "Brain" system. It manages the full lifecycle of AI-driven development—from task delegation and state synchronization to autonomous verification and fleet-wide orchestration.
JCLAW operates on the Lobster Pattern, cleanly separating the AI's reasoning from its execution environment:
- 🧠 The Brain (Reasoning): Powered by the JCLAW Dispatcher and Gemini LLMs. It analyzes project DNA, identifies risks, provides technical feedback to blocked tasks, and performs final verification of finished work.
- 🦞 The Muscles (Execution): Powered by Google Jules. It performs the heavy lifting—refactoring code, writing tests, and fulfilling instructions in an isolated remote VM.
- 🌊 The Tide (Orchestration): Powered by Superclaw and Autopilot. It synchronizes state across the fleet, manages git worktrees for zero-conflict execution, and updates GitHub Project boards in real-time.
Initialize your environment variables and authenticate with the Jules API:
jclaw setupRequired ENV: JULES_API_KEY, JCLAW_BRAIN_AGENT (default: gemini).
Instead of typing commands, you can manage your work through the Backlog:
- Create a markdown task in
.jules/backlog/feature-xyz.md. - Run the orchestrator:
jclaw wave --superclaw
JCLAW will detect the workload, scaffold an isolated git worktree, delegate the task to Jules, and notify you when the Brain has performed the final audit.
superclaw.sh: The high-level fleet manager. It iterates through all repositories in your workspace, detects workloads, and initializes Lazy Git Worktrees to run automation without disrupting your active IDE buffer.autopilot.sh: The heart of the per-repo cycle. It handles:- Auto-Approval: Automatically approves safe execution plans.
- Verification Gates: Dispatches finished code to the Brain for audit.
- Feedback Loops: Answers Jules' questions autonomously using LLM context.
- Batching: Pushes multiple backlog items concurrently within API limits.
sync.sh: Bilateral synchronization. It pulls remote session status into the local SQLite DB and pushes state updates to the GitHub Project Board and a dedicatedjclaw-statebranch.
brain_dispatcher.sh: Routes "Brain Requests" (Verification/Feedback) to your selected AI agent (Gemini, Cursor, or browser-based LLMs).brain_context_builder.sh: The "High-Fidelity" context engine. It assembles Project DNA, file churn risk profiles, and active wave dependencies to ensure the AI has perfect technical awareness.brain_report_parser.sh: Parses AI responses into Markdown audit reports and automatically posts them as PR comments.
delegate_task.sh: Initiates a specific local goal into a remote Jules session.conclude_task.sh: Finalizes a task, archives the instruction file, and generates the final conclusion logs.github_integration.sh: The bridge to GitHub. Manages Issue/PR detection, labeling, and Project Board automation.db_manager.sh: Manages the local SQLite cache (.jules/jclaw.db) for high-speed session tracking.
JCLAW solves the "Git Contention" problem by never running destructive operations in your primary working directory. When superclaw detects a workload:
- It creates a temporary (or persistent) Git Worktree at
[repo]-jclaw/. - It executes all stashes, checkouts, and merges inside that isolated environment.
- Your own code, index, and branch state remain completely untouched.
JCLAW maintains a live dashboard for your project:
- Local: Check
.jules/README.mdfor a generated state overview. - Remote: Check the
jclaw-statebranch on GitHub for synchronized JSON metadata. - Board: View progress in real-time on your configured GitHub Project Board.
Developed by the Advanced Agentic Coding team.