Mirage is a Unified Virtual File System for AI Agents: a single tree that mounts services and data sources like S3, Google Drive, Slack, Gmail, and Redis side-by-side as one filesystem.

AI agents reach every backend with the same handful of Unix-like tools, and pipelines compose across services as naturally as on a local disk. It's a simulated environment, agents see one filesystem underneath. Any LLM that already knows bash can use Mirage out of the box, with zero new vocabulary.

constws=newWorkspace({'/data': newRAMResource(),'/s3': newS3Resource({bucket: 'logs'}),'/slack': newSlackResource({}),'/github': newGitHubResource({}),})awaitws.execute('grep alert /slack/general/*.json | wc -l')awaitws.execute('cat /github/mirage/README.md')awaitws.execute('cp /s3/report.csv /data/local.csv')// Register a new command, available across every mount.ws.command('summarize', ...)// Override a command for a specific resource + filetype —// `cat` on a Parquet file in /s3 renders rows as JSON instead of raw bytes.ws.command('cat',{resource: 's3',filetype: 'parquet'}, ...)awaitws.execute('summarize /github/mirage/README.md')awaitws.execute('cat /s3/events/2026-05-06.parquet | jq .user')

About

• One filesystem, every backend. Every service speaks the same filesystem semantics, so agents reason about one abstraction instead of N SDKs and M MCPs, leaning on the filesystem and bash vocabulary LLMs are most fluent in.
• Multiple resources, one filesystem: RAM, Disk, Redis, S3 / R2 / OCI / Supabase / GCS, Gmail / GDrive / GDocs / GSheets / GSlides, GitHub / Linear / Notion / Trello, Slack / Discord / Telegram / Email, MongoDB, SSH, and more, mounted side-by-side under a single root.
• Familiar bash tools across every mount. Agents reuse the same handful of Unix-like tools instead of learning a new API per service, and pipelines compose across services as naturally as on a local disk, the exact corpus modern LLMs are most heavily trained on.
• Portable workspaces: clone, snapshot, and version your environment. Move agent runs between machines without restarting or reconfiguring the system.
• Embed in your apps and services: Python and TypeScript SDKs let you give your AI agents a virtual filesystem directly inside FastAPI, Express, browser apps, or any async runtime, no separate process required. Clone, snapshot, and version the workspace from inside your code.
• Works with major agent application frameworks: OpenAI Agents SDK, Vercel AI SDK (TypeScript), LangChain, Pydantic AI, CAMEL, and OpenHands.
• Lightweight CLI + daemon: plugs into coding agents like Claude Code and Codex so they reach every mounted resource through familiar bash, getting more useful work done per turn.

Architecture

Installation

Prerequisites

• Python ≥ 3.12 for the mirage-ai package and the mirage CLI
• Node.js ≥ 20 for the TypeScript SDK
• macOS or Linux (FUSE-based mounts require platform support)

Python

uv add mirage-ai

This installs both the mirage library and the mirage CLI binary.

TypeScript

Pick the package that matches your runtime:

npm install @struktoai/mirage-node # Node.js servers and CLIs
npm install @struktoai/mirage-browser # browser / edge runtimes
npm install @struktoai/mirage-core # runtime-agnostic primitives

@struktoai/mirage-node and @struktoai/mirage-browser both pull in @struktoai/mirage-core automatically.

CLI

curl -fsSL https://strukto.ai/mirage/install.sh | sh

Or via your package manager of choice:

npm install -g @struktoai/mirage-cli

uvx mirage-ai

npx @struktoai/mirage-cli

Quickstart (Python)

frommirageimportWorkspacefrommirage.resource.gdocsimportGDocsConfig, GDocsResourcefrommirage.resource.ramimportRAMResourcefrommirage.resource.s3importS3Config, S3Resourcefrommirage.resource.slackimportSlackConfig, SlackResourcews=Workspace({
"/data": RAMResource(),
"/s3": S3Resource(S3Config(bucket="my-bucket")),
"/slack": SlackResource(SlackConfig()),
"/docs": GDocsResource(GDocsConfig()),
})
awaitws.execute("cp /s3/report.csv /data/report.csv")
awaitws.execute("grep alert /s3/data/log.jsonl | wc -l")
ws.snapshot("demo.tar")

Quickstart (TypeScript)

import{Workspace,RAMResource,S3Resource,SlackResource,GDocsResource,}from'@struktoai/mirage-browser'constws=newWorkspace({'/data': newRAMResource(),'/s3': newS3Resource({bucket: 'my-bucket'}),'/slack': newSlackResource({}),'/docs': newGDocsResource({}),})awaitws.execute('cp /s3/report.csv /data/report.csv')awaitws.execute('grep alert /s3/data/log.jsonl | wc -l')

Quickstart (CLI)

mirage workspace create ws.yaml --id demo
mirage execute --workspace_id demo --command "cp /s3/report.csv /data/report.csv"
mirage provision --workspace_id demo --command "cat /s3/data/large.jsonl"
mirage workspace snapshot demo demo.tar
mirage workspace load demo.tar --id demo-restored

Agent Frameworks

Mirage drops into the major agent application frameworks as a sandbox or tool layer. Your agent runs against the same mount tree it would in bash, so swapping the model or runtime never changes the surface.

OpenAI Agents SDK (Python)

The MirageSandboxClient plugs a Workspace into the OpenAI Agents SDK as a sandbox: bash commands the agent runs execute against your mounts.

fromagentsimportRunnerfromagents.runimportRunConfigfromagents.sandboximportSandboxAgent, SandboxRunConfigfrommirage.agents.openai_agentsimportMirageSandboxClientclient=MirageSandboxClient(ws)
agent=SandboxAgent(
name="Mirage Sandbox Agent",
model="gpt-5.4-nano",
instructions=ws.file_prompt,
)
result=awaitRunner.run(
agent,
"Summarize /s3/data/report.parquet into /report.txt.",
run_config=RunConfig(sandbox=SandboxRunConfig(client=client)),
)

Vercel AI SDK (TypeScript)

mirageTools(ws) exposes the workspace as a typed AI SDK tool set, so any model wired into the AI SDK can read and write across mounts, in Node or the browser.

import{generateText}from'ai'import{openai}from'@ai-sdk/openai'import{mirageTools}from'@struktoai/mirage-agents/vercel'import{buildSystemPrompt}from'@struktoai/mirage-agents/openai'const{ text }=awaitgenerateText({model: openai('gpt-5.4-nano'),system: buildSystemPrompt({mountInfo: {'/': 'In-memory filesystem'}}),prompt: "Use readFile to read /docs/paper.pdf, then describe what's in it.",tools: mirageTools(ws),})

LangChain, Pydantic AI, CAMEL, OpenHands, and Mastra adapters live alongside these.

Cache

Every Workspace ships with a two-layer cache so repeated work against remote backends (S3, GDrive, Slack, …) hits local state instead of the network:

• Index cache. Listings and metadata. The first directory walk hits the API; subsequent ones serve from the index until TTL expires.
• File cache. Object bytes. The first read streams from origin; later pipelines read from cache.
• Pluggable backends. Each layer is a store with two built-ins:
  • RAM (default): in-process, zero setup, 512 MB file cache and 10-minute index TTL. Best for single-process apps and notebooks.
  • Redis: shared across workers, processes, and machines. Best for serverless, multi-replica services, or when you want cache state to survive restarts.

import{RedisFileCacheStore,RedisIndexCacheStore,Workspace}from'mirage/node'constws=newWorkspace({'/s3': newS3Resource({bucket: 'my-bucket'})},{cache: newRedisFileCacheStore({url: 'redis://localhost:6379/0',limit: '8GB'}),index: newRedisIndexCacheStore({url: 'redis://localhost:6379/0',ttl: 600}),},)

import{S3Resource,Workspace}from'mirage/node'constws=newWorkspace({'/s3': newS3Resource({bucket: 'my-bucket'})})// 1. Index miss → S3 LIST. Listing stored in index cache.awaitws.execute('ls /s3/data/')// 2. Index hit → 0 network calls.awaitws.execute('find /s3/data/ -name "*.jsonl"')// 3. File miss → S3 GET. Bytes stored in file cache.awaitws.execute('cat /s3/data/log.jsonl | wc -l')// 4. File hit → 0 network calls.awaitws.execute('grep alert /s3/data/log.jsonl')