🌐 Mirage：2.1k Stars 的统一虚拟文件系统，让 AI Agent 用 bash 操作一切后端

项目地址：github.com/strukto-ai/mirage | ⭐ 2,135 Stars | 🛠 TypeScript/Python | 📅 2026-05-06

老实说，现在的 AI Agent 生态有个挺尴尬的事——每个后端服务都要学一套新 API。S3 一套 SDK，Slack 一套，Gmail 又一套，再加上 MCP server 满天飞。Agent 的 tool-use 调用链路越搞越长，debug 起来想骂人。

Mirage 的思路就很暴力：把所有后端挂成一个虚拟文件系统，Agent 直接用 bash 操作。cp、grep、cat、jq——LLM 最熟悉的那些 Unix 命令，直接操作 S3、Slack、GitHub、Gmail 的数据。不需要学新东西。

一、核心思路：一切皆文件

Mirage 把每个服务抽象成一个 "Resource"，全部挂到同一个目录树下：

from mirage import Workspace
from mirage.resource.s3 import S3Config, S3Resource
from mirage.resource.slack import SlackResource

ws = Workspace({
    "/data":  RAMResource(),
    "/s3":    S3Resource(S3Config(bucket="my-bucket")),
    "/slack": SlackResource(),
    "/docs":  GDocsResource(),
})

配完这些，你的 Agent 就能在同一个 bash session 里跨服务操作：

# 跨服务管道——从 Slack 捞告警，跟 S3 数据做关联
grep alert /slack/general/*.json | wc -l
cp /s3/report.csv /data/report.csv
cat /docs/mirage/README.md

最骚的操作是 cp 在跨服务间直接工作——/s3/report.csv 复制到 /data/ 就是 RAM 里的本地文件，Agent 不需要知道底层在干嘛。

二、快速上手

装一个 CLI 就行：

# Python
uv add mirage-ai

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

# TypeScript
npm install @struktoai/mirage-node

创建 workspace 并跑命令：

mirage workspace create ws.yaml --id demo
mirage execute --workspace_id demo --command "cp /s3/report.csv /data/report.csv"
mirage workspace snapshot demo demo.tar

三、跟 Agent 框架集成

Mirage 原生支持 OpenAI Agents SDK、LangChain、Vercel AI SDK 等主流框架。最简单的玩法是直接给 Claude Code 挂上虚拟文件系统：

from agents import Runner
from agents.sandbox import SandboxAgent
from mirage.agents.openai_agents import MirageSandboxClient

client = MirageSandboxClient(ws)
agent = SandboxAgent(
    name="Mirage Agent",
    instructions=ws.file_prompt,
)
result = await Runner.run(agent, "Summarize /s3/data/logs.jsonl")

TypeScript 也能用：

import { Workspace, S3Resource, SlackResource } from '@struktoai/mirage-browser'

const ws = new Workspace({
  '/s3':    new S3Resource({ bucket: 'my-bucket' }),
  '/slack': new SlackResource({}),
})

await ws.execute('grep alert /slack/general/*.json | wc -l')

四、跟 MCP 对比

Mirage 不是 MCP 的替代品，思路更接近 FUSE 文件系统。MCP 给 Agent 暴露一个个独立的 tool endpoint，Mirage 直接给一个统一的文件系统抽象。好处是：

五、目前能挂什么

官方支持的 Resource 列表：S3/R2/GCS、Gmail/GDrive/GDocs、GitHub/Linear/Notion、Slack/Discord/Telegram、Redis、MongoDB、SSH 远程主机，还有 RAM 和本地磁盘。

总结

🌐 Mirage: 2.1k Stars Unified Virtual Filesystem — Your AI Agent Uses Bash for Everything

GitHub: github.com/strukto-ai/mirage | ⭐ 2,135 Stars | 🛠 TypeScript/Python

Let's be real — the current AI agent ecosystem has a friction problem. Every backend service needs a different SDK, a different tool definition, and sometimes even a different MCP server. The function-calling chain gets absurdly long.

Mirage takes a different approach: mount every backend as a virtual filesystem, and let your agent use plain Unix commands. cp, grep, cat, jq — the commands LLMs know best — work across S3, Slack, GitHub, and Gmail out of the box. No new vocabulary needed.

Core Idea: Everything is a File

Mirage abstracts each service as a "Resource" and mounts them under one directory tree:

from mirage import Workspace
from mirage.resource.s3 import S3Config, S3Resource
from mirage.resource.slack import SlackResource

ws = Workspace({
    "/data":  RAMResource(),
    "/s3":    S3Resource(S3Config(bucket="my-bucket")),
    "/slack": SlackResource(),
    "/docs":  GDocsResource(),
})

Once configured, your agent can pipe data across services in a single bash session:

grep alert /slack/general/*.json | wc -l
cp /s3/report.csv /data/report.csv
cat /docs/mirage/README.md

The killer feature: cp works cross-service. Copying from /s3/report.csv to /data/ places it into RAM — the agent doesn't need to know what happens underneath.

Quick Start

# Python
uv add mirage-ai

# CLI one-liner
curl -fsSL https://strukto.ai/mirage/install.sh | sh

# TypeScript
npm install @struktoai/mirage-node

Create a workspace and run commands:

mirage workspace create ws.yaml --id demo
mirage execute --workspace_id demo --command "cp /s3/report.csv /data/report.csv"
mirage workspace snapshot demo demo.tar

Agent Framework Integration

Mirage drops into OpenAI Agents SDK, LangChain, and Vercel AI SDK as a sandbox layer:

from agents import Runner
from agents.sandbox import SandboxAgent
from mirage.agents.openai_agents import MirageSandboxClient

client = MirageSandboxClient(ws)
agent = SandboxAgent(name="Mirage Agent", instructions=ws.file_prompt)
result = await Runner.run(agent, "Summarize /s3/data/logs.jsonl")

TypeScript equivalent:

import { Workspace, S3Resource, SlackResource } from '@struktoai/mirage-browser'

const ws = new Workspace({
  '/s3': new S3Resource({ bucket: 'my-bucket' }),
  '/slack': new SlackResource({}),
})
await ws.execute('grep alert /slack/general/*.json | wc -l')

Mirage vs MCP

Mirage isn't a replacement for MCP — it takes a FUSE-like approach. MCP exposes individual tool endpoints; Mirage gives a unified filesystem abstraction. The wins are:

Supported Resources

S3/R2/GCS, Gmail/GDrive/GDocs, GitHub/Linear/Notion, Slack/Discord/Telegram, Redis, MongoDB, SSH hosts, RAM, and local disk — mounted side by side under one root.

TL;DR