The Filesystem for Agents (github.com)

AgentFS is an alpha release of a filesystem designed specifically for AI agents: a single, SQLite-based runtime (built on Turso) that unifies files, agent state, and an audit trail into one portable database file. It exposes a POSIX-like virtual filesystem for file I/O, a key-value store for agent state/context, and tooling hooks that log every tool invocation. Ship-ready primitives include a CLI (agentfs init / run / fs ...), TypeScript and Rust SDKs (examples: await agent.kv.set('user:preferences', {...}); await agent.fs.writeFile('/output/report.pdf', buffer); await agent.tools.record('web_search', ...)), and a Linux-compatible sandbox to run agents with controlled filesystem access. The project is MIT-licensed and explicitly alpha—intended for development and experimentation only. For the AI/ML community this matters because it makes agent behavior auditable, reproducible, and portable: every file operation, state change, and tool call is queryable via SQL, and you can snapshot/restore a complete agent runtime with cp agent.db snapshot.db. That simplifies debugging, compliance, and “what-if” testing, and enables versioning or CI workflows by treating the whole agent runtime as a single artifact. It removes dependency on external DB servers and distributed storage for many workflows, but users should avoid production-critical data until maturity and security hardening are proven.