Mintlify had a problem. Their AI documentation assistant used sandboxes that cloned repositories to give agents real filesystem access, but session creation took 46 seconds at p90. Users stared at loading spinners. At 850,000 conversations a month, the infrastructure bill would have topped $70,000 annually. So they built ChromaFs, a virtual filesystem that fakes it. Instead of spinning up containers, ChromaFs intercepts UNIX commands like grep, cat, ls, and cd, then translates them into queries against their existing Chroma database. Session creation dropped to roughly 100 milliseconds. Marginal compute cost hit zero because they're reusing infrastructure they already pay for.
The technical approach is clever. ChromaFs runs on just-bash, Vercel Labs' TypeScript reimplementation of bash with a pluggable filesystem interface. When the agent boots, ChromaFs fetches a gzipped JSON document containing the entire file tree and loads it into memory. Operations like ls and find resolve locally with no network calls. When the agent runs cat or grep, ChromaFs translates those into targeted database queries. For grep specifically, it uses Chroma as a coarse filter to identify matching files, prefetches those chunks into Redis, then hands off to just-bash for in-memory regex execution. The whole thing is read-only, so there is no state to clean up between sessions.
Role-based access control happens before the agent sees anything. The file tree includes metadata about which user groups can access which paths. ChromaFs prunes inaccessible files before building the in-memory structures, so the agent cannot even reference paths the user lacks permission to see. What would require managing Linux user groups and container images per customer tier becomes a few lines of filtering code. The approach has sparked debate about whether grep-based search is really an improvement over vector-based RAG or just a return to keyword search. But the performance numbers speak for themselves, and filesystem interfaces are clearly becoming a default interaction model for AI agents.