Mirage: The Universal File System for AI Agents

How one open-source project gives AI agents a single root directory across every platform, database, and cloud service.

The problem with AI agents today is not intelligence. It is plumbing.

Every platform your agent might need lives in its own world. S3 has one SDK. Google Drive has another. Slack has a web API. GitHub has its own REST interface. Linear, Notion, and Postgres each require separate authentication, separate client libraries, and separate mental models. For a human engineer, this is manageable. For an AI agent trying to act autonomously across all of them, it is an endless series of translation problems.

Strukto AI published Mirage to fix this at the root.

The core idea is a virtual filesystem. Mirage mounts every backend — cloud storage, databases, SaaS platforms — under a single directory tree. Your agent does not need to learn that S3 uses `boto3` and Google Drive uses OAuth scopes. It just navigates to `/s3/my-bucket/` or `/drive/reports/` and runs the same commands it already knows.

That is the second insight: zero new vocabulary. The team built Mirage around standard bash commands. `cat`, `grep`, `head`, `ls`. Every language model was trained on millions of examples of these commands. Mirage makes them work natively on `.parquet`, `.csv`, `.json`, and `.wav` files without any preprocessing step. To make that happen, the team rewrote bash from scratch in over 1.1 million lines of code.

Because everything is unified under one namespace, pipelines become trivial to compose. A single shell pipe can pull a file from S3, filter it with `grep`, join it against a Postgres query, and push the result to a Slack channel. No glue code. No intermediate format conversions. To keep repeated reads fast, Mirage runs a two-layer cache that stores hot data locally, so agents are not hammering remote APIs on every access.

The feature that may matter most in practice is versioning. Mirage lets you snapshot an entire multi-platform workspace — every mounted backend, every file, every database state — and store it as a point-in-time checkpoint. You can clone that snapshot, branch from it, or roll back to it with a single API call. For agentic workflows where mistakes are expensive and reproducibility matters, this is the equivalent of `git` for your agent's entire environment.

Mirage is designed to embed directly inside the frameworks agents already run in. It ships as a native module for FastAPI and Express, and integrates with the OpenAI Agents SDK, Vercel AI SDK, LangChain, Claude Code, and Codex. You do not retool your infrastructure to use it. You drop it in.

It is entirely open source.

The deeper bet Mirage is making is that the bottleneck for agentic AI is not model capability — it is the friction of acting in the real world. Every silo an agent has to learn costs tokens, latency, and failure surface. A universal filesystem does not just make things easier. It changes what is possible.