emotions-research/lunarfs

LunarFS is a filesystem tool that lets you create isolated, disposable copies of a code repository in milliseconds rather than seconds. The core idea is that instead of physically copying every file when you want a new workspace, it copies only a single 32-byte identifier that represents the entire repository as a content-addressed snapshot. The actual files are then loaded on demand the first time each one is read. This approach makes it practical to run many separate workspaces over the same repository at the same time, which is especially useful when running multiple AI agents in parallel. Each agent gets its own isolated copy that starts almost instantly, each can make changes independently, and when you are done you can inspect exactly what each agent changed and whether any two agents modified the same file. A benchmarked example shows forking the Linux kernel taking 13 milliseconds with LunarFS compared to 7.4 seconds with the git worktree equivalent. The filesystem is built in Rust and uses BLAKE3, a cryptographic hashing algorithm, to identify and deduplicate every file blob. Mounting a workspace uses the operating system's built-in NFS client on macOS and Linux, so no additional drivers are required. An MCP server, which is a protocol some AI coding tools use to extend their capabilities, lets AI agents create and destroy workspaces, list what they have access to, and push revisions without leaving the agent interface. The tool also includes a sync feature that watches a directory and automatically pushes changes across machines without a manual step. A hosted service at lunarfs.com handles the server side for teams that do not want to run their own. The engine is licensed under AGPL-3.0, which means you can use it freely but must share modifications. The client libraries are licensed under Apache-2.0, which is more permissive.