sanjoydat1/p2pa

P2PA is a local-first toolkit that lets multiple AI coding assistants share context with each other directly, without routing through a central cloud service. It is designed for founders and engineering teams who run several AI agents and want those agents to pass information, divide up work, and merge edits without someone manually copying prompts between them. The core problem it addresses is that local AI agents tend to work in isolation. When one finishes a task, its working memory disappears, so another agent on a different machine starts from scratch. Passing full context windows through a central server burns tokens and adds delay. P2PA keeps a shared state file on each machine and synchronizes only the changes. It uses a peer-to-peer network layer called Hyperswarm, so there is no central server to stand up. Peers authenticate each other using cryptographic keys, and only keys you have explicitly allowed can connect. Knowing the network topic is not enough. The merge logic is built around conflict-free data types. Two agents writing to different keys never conflict. Two agents writing to the same key resolve to the same result on every machine, with no central arbiter. There are work-claiming leases so two agents never duplicate the same task, and those leases expire automatically if an agent crashes. Messages written to an offline peer are queued and delivered when that peer returns. Every change is written to a plain Markdown audit log on disk, attributed to the peer that made it, so a human can read the full session without any tools. You can test it on a single machine with a few npm scripts that exercise the merge engine, the work-claiming system, and the offline message queue. The test suite runs fully offline. To use it between two machines, you install the package globally, run a pairing command on each side to exchange keys, and then connect an IDE agent like Cursor or Claude Code through a generated config block. You choose one of two process modes: a foreground MCP mode for when an IDE agent is driving, or a background daemon mode for standalone sync. Both modes write the same files, and a writer lock prevents them from running at the same time. The README includes a full CLI reference, a walkthrough of a sample two-agent session, and a Google Docs integration for shared war rooms. The full README is longer than what was shown.