Skip to content

Introduction

What AFI is

The Agent Filesystem Interface (AFI) is a specification for exposing arbitrary data sources to AI agents as a POSIX-shaped filesystem. Any system that implements AFI is an AFI provider. Any agent runtime that consumes AFI providers is an AFI client.

The interface is small on purpose. Four core operations — stat, list, read, search — plus discovery, access control, and four transport bindings (MCP, HTTP, FUSE, SSH).

What AFI is not

  • Not a wire protocol. Any of the transport bindings can be conformant.
  • Not a specific implementation. SQLite, vector DBs, object stores, HTTP APIs all work as backing stores.
  • Not a replacement for MCP. AFI ships an MCP binding as the recommended transport — the two compose.
  • Not a replacement for POSIX or FUSE. AFI borrows their mental model.

The problem

LLM agents are trained on massive volumes of code. Their default reasoning about "how do I explore an unfamiliar body of information?" is a bash session: ls to see the shape, find to locate something specific, grep to search across, cat to read.

Recognizing this, several teams shipped virtual-filesystem interfaces over their data in 2025–2026 — Mintlify over their docs index, Supabase over their docs corpus via SSH, Turso over agent state via FUSE, Vercel over their analytics DSL via a bash tool. Each solution works. None of them share a shape.

That means:

  • Agent runtimes must integrate each provider one-off
  • A skill learned against one provider doesn't transfer to another
  • Every new provider re-invents the same operations, with subtle differences
  • No conformance testing means "does it actually work like a filesystem?" is an implementer judgment call

AFI is a small standard that fixes exactly these problems, and nothing else.

The AFI stance

  • A tight core, generous extensions. Four required operations, structured extension points (x- prefixed fields), an AEP process for standardizing what the ecosystem converges on.
  • Transport-plural. MCP is recommended, HTTP is offered for direct integrations, FUSE serves the "make it look like a real disk" use case, SSH serves the "any developer with a terminal can poke at it" use case.
  • Conformance-tested. Four profiles from discovery-only to full read/write; providers declare their level and a public test kit checks that declaration is honest.
  • Vendor-neutral from day one. The governance model requires two independent implementations before v1.0, and the project aims for a neutral home once that gate is met.

Where to go next