Introduction

AgentSync keeps a swarm of coding agents from breaking your repo. It does two things: it autonomously maintains the skill files that keep agents fluent in your codebase, and it coordinates every agent working in a repo — real-time presence, claim arbitration, and conflict prediction — so overlapping work is denied before it starts, not discovered at merge time.

The core idea

One coding agent is easy to manage. Ten or thirty running in parallel on the same repo is chaos: agents edit the same files, branches go stale against a moving main, work gets duplicated, and pull requests collide at merge time — after the work is already done. Git tells you what changed, after the fact. It does not tell you who is about to touch what.

AgentSync sits between your agents and git and answers the questions git cannot:

• Who is working on what, right now.
• Which files each agent intends to touch.
• Which branches are stale against the default branch.
• Which agents are on a collision course, before they collide.
• What order changes should merge in.

npm i -g @remotehostai/agentsync
agentsync init

It runs standalone on a laptop with zero dependency on any hosted service — plain git, no account required — or against a self-hosted or RemoteHost-hosted coordinator for the full deterministic guarantee.

What's in these docs

• Getting Started — install the CLI and run your first claim.
• Concepts — actors, tasks, claims, and the coordinator.
• CLI Reference — every agentsync command and what it does.
• Protocol — the AgentSync Protocol (ASP): invariants and the config.yaml schema.
• Self-Hosting — run the reference coordinator on your own Cloudflare account.