An AI task board that ships code via git worktrees and multi-node sync

What it does

Fusion is a multi-node orchestrator that takes a plain-language task, has a planning agent write a structured PROMPT.md with steps and acceptance criteria, then drives the work through plan-review-execute-review gates until it squash-merges into your repo. It runs across desktop, mobile, web, and CLI nodes that stay synchronized, and isolates every task in its own git worktree to keep parallel work from stepping on each other.

The interesting bit

The project treats git worktrees as first-class execution sandboxes—each task gets a fusion/{task-id} branch and worktree, enabling parallel agents without file conflicts. It also borrows the companies.sh standard to let you import pre-built “agent companies” (440+ agents across 16 teams) that coordinate via internal mailboxes and hierarchical missions for weeks of autonomous work.

Key highlights

• AI planning with explicit PROMPT.md specs, scoped files, and acceptance criteria before any code is written
• Workflow gates between every step: plan, review, execute, review again, with optional human approval
• Multi-node mesh syncs state across laptops, servers, phones, and cloud VMs via desktop, mobile, web, and CLI clients
• Worktree isolation means parallel tasks run concurrently on zero-conflict branches
• Supports local and cloud LLMs (Anthropic, OpenAI, Ollama, etc.) and experimental plugins for Hermes, Paperclip, and OpenClaw runtimes

Caveats

• Explicitly labeled as “early preview” with weekly releases; expect rough edges and shifting APIs
• Several major features—including Multi-agent Chat Rooms and the Hermes, Paperclip, and OpenClaw plugins—are experimental, with wire formats that may change between minor releases
• Tasks with dependencies are forced to run sequentially, so only independent tasks get the full benefit of parallel worktree isolation

Verdict

Worth a look if you want an open-source, self-hostable alternative to managed AI coding agents with strong git hygiene and cross-device control. Skip it if you need battle-tested stability or a simple, single-machine script runner—the mesh and orchestration layers are still baking.