The versioned filesystem for AI agents

Branch-based versioning is a huge unlock for agent applications. If you want an agent to do work that is then reviewed for approval by humans, or if you want to run parallel agent swarms that modify the same files, or if you want have an audit trail for your agents, or support checkpoints to rollback mistakes then you probably want versioning.

You can! You'll find that you run into rate limits, repo limits, and ergonomics that make you want to tear your hair out, though. Git providers give you branching, history, and diffs but no real filesystem — you can't mount them, random-access large files, or get sub-50ms reads. Mesa gives you all of Git's versioning semantics plus a fast, mountable filesystem in one platform.

If you just need a durable filesystem and you're on AWS then S3 Files is a great solution. S3 Files only works on AWS services, however, so if you anticipate your agent running in multiple environments you'll need a cross-cloud solution. Similarly, S3 Files doesn't support branch-basedversioning or diffs which means if you want agent swarms to edit documents in parallel you'll need to implement some level of locking or just let the agents clobber over each other. Mesa gives you built-in primitives to scale to as many agents as you need without worrying about locking.

Yes. MesaFS is 100% POSIX compatible so agents and sandboxes can use familiar Unix tools and standard filesystem APIs. Mesa handles persistence, sync, and version history transparently underneath.

Mesa has built-in support for syncing data between Git upstreams like GitHub, GitLab, and others. We aim to support non-Git upstreams like S3, Google Drive, and others in the future but for now it's easy to write your own sync scripts using our APIs and MesaFS.

You can read data from Mesa in 3 ways: Git, our REST API's content endpoint, and our virtual filesystem (MesaFS) which can be mounted in your application or on the OS itself. We meter per GB of egress for Git and our Content API but all reads through MesaFS are free and do not count towards your egress limit. MesaFS is the primary way to interact with repositories in Mesa.

Yes. Mesa supports large files with random access, no per-file size limits, and built-in deduplication. Store code, datasets, models, and media in the same workspace.

Sub-50ms reads and writes through MesaFS, p95 reads under 100ms through our API, and millisecond-level forks and branches. Spinning up an isolated workspace for a parallel agent run is essentially free.

When building agent applications, performance generally comes down to time-to-first-token (TTFT) and read/write latency. Mesa's TTFT is lower than Git by an order of magnitude because you avoid the cost of cloning an entire repo. Mesa materializes files on demandso your agent can start working immediately. Mesa also maintains a local cache and intelligently prefetches data based on access patterns. For warm reads, Mesa's p95 latency is on the order of nanoseconds and for cold reads it's under 100ms.