We're Dari (YC F25). We're building browser automation agents with a deterministic caching layer; for that to work, we needed a way to save and rerun code across sessions. There's also no audit trail with traditional sandboxes; when something goes wrong, you have no idea what the agent actually did.

So we built SkillFS. The core idea is simple: every agent sandbox is a git repo. The agent does its work, commits its progress, and the session ends. When you start a new session, SkillFS restores the repo from a git bundle and the agent continues where it left off. And because it's git, you have a complete history of every action the agent took.

The workflow looks like this:

  1. Agent works in a sandboxed environment, making changes to files
  2. Agent commits progress at meaningful checkpoints
  3. Session ends, git bundle gets saved to storage (local or GCS)
  4. Next session starts, bundle is restored, agent resumes
  5. Need to debug? git log shows you exactly what happened

Key features:

  - Persistent state via git bundles, with pluggable storage backends (local filesystem or GCS), so skills and scripts from previous sessions are reusable.

  - MCP integration that lets you plug in any server. SkillFS also generates Python wrappers from MCP tool definitions and uploads them to the sandbox, so your agent can call any MCP tool as regular code to cache sequences of MCP interactions into deterministic scripts.

  - Built-in LLM runner with standard tools (glob, grep, read/write/edit files, run commands), or you can bring your own agent loop.

  - Runs on E2B sandboxes so agents execute code in isolated environments. Agent skills can also be easily imported from local files or from GitHub repos.

We're open sourcing this because agent persistence is a problem every team building with agents ends up solving differently, and we think bash+git is a good answer that more people should be using.

  > pip install skillfs