How We Built a Virtual Filesystem for Our Assistant — Mintlify

URL: https://www.mintlify.com/blog/how-we-built-a-virtual-filesystem-for-our-assistant
Author: Dens Sumesh (Mintlify Engineering)
Published: March 24, 2026

Summary

Mintlify built ChromaFs — a virtual filesystem backed by their existing Chroma database — to let their documentation assistant explore docs using shell commands (grep, cat, ls, find) without spinning up real containers. Session creation time dropped from 46 seconds to 100 milliseconds, and marginal compute cost dropped to ~$0.

Key Claims

  • Agents converge on filesystems: grep, cat, ls, find are the natural interface for agent document exploration. If each doc page is a file, agents can navigate exactly like exploring a codebase.
  • Real containers are too slow for interactive use: Sandbox p90 boot time was ~46 seconds. Cloning GitHub repos added further latency. At 850K conversations/month, cost would exceed $70K/year.
  • ChromaFs = virtual shell on Chroma: Intercepts UNIX commands and translates them into Chroma queries against the already-indexed documentation. Built on just-bash (Vercel Labs TypeScript bash reimplementation).
  • Path tree trick: A gzipped JSON document __path_tree__ inside Chroma stores the entire file tree. On init, ChromaFs decompresses it into in-memory Set/Map structures. ls, cd, find resolve locally — no network calls.
  • Access control as filter: isPublic and groups fields in the path tree let ChromaFs prune paths before building the tree. Per-user access control without Linux user groups or container images.
  • Performance: 46s → 100ms boot; $0 marginal cost (reuses existing DB).

Implications for Context Engineering

This is a concrete implementation of the “filesystem as agent memory” pattern. Rather than static retrieval, the agent can explore, grep for exact strings, and navigate hierarchically. Compare to Context Repositories — Letta which takes a git-backed approach to persistent agent memory.