Unified Interface
One virtual filesystem presents every service as files and commands, and agents read them live without copying anything locally.
AgentVFS
The Virtual Filesystem Interface for AI Agents
Why AgentVFS?
Performance
Agents filter and chain data with Unix-like tools and scripts, saving more than half of MCP's tokens on average with better accuracy.
Policy & Security
The filesystem is purely virtual, so a policy checks every command and decides what agents may read, change or run, keeping secrets and tests out of reach.
Inside the benchmark runs.
Three moments from VFS-Bench 1.0, in the agents' own commands.
Unified Interface
In canvas-art-manager, Canvas and the mailbox are folders in one virtual filesystem, one interface to both, so agents read the course schedule straight from its email with the same commands they use on Canvas.
- Resources
- 47
- Tasks
- 200
- Virtual Filesystem
- 1
Performance
On the same task, GPT-5.6 Luna creates one course to try the command, then the other 26 in one loop with VFS, while with MCP it makes a tool call per step until it runs out of requests.
MCPRequests 0Tokens 0.00M
VFSRequests 0Tokens 0.00M
root~
- Token reduction
- Up to 88%
- Average success rate increase
- +5.5%
Policy & Security
Across Gemini 3.8 Flash's Toolathlon runs, the policy refuses stray commands, such as searching the whole virtual filesystem for tokens, as in sync-todo-to-readme, or for the benchmark itself, guarding against leaks and reward hacking.
- Runs with a stray command refused
- 74 of 100
- Runs with a costly search from /
- 40 of 100
- Policy
- 1
Our Mission
We explore how AI agents should work with data storage, cloud services and the applications people use every day, to find the interface that gives them the best performance at the least cost. Along the way, we study where today's models fall short with these systems and how they can improve. AgentVFS is an open, vendor-neutral specification and benchmark that any implementation and any model can use, and we develop it in the open.