The World's First Virtual Terminal for AI Agents

One unified interface for AI agents, with a virtual filesystem and swappable runtimes across machines, services, and clouds. Scale your agents with fewer tokens and less infrastructure.

Read the docs 3.6K

npm install @struktoai/mirage-node
npm install @struktoai/mirage-browser

const ws = new Workspace({
  '/tmp':   [new RAMResource(), MountMode.EXEC],
  '/redis': [new RedisResource({ url: redisUrl }), MountMode.WRITE],
  '/slack': [new SlackResource({ token: slackToken }), MountMode.EXEC],
}, { runtimes: [buildRuntime('monty', { captures: ['python'] }), 'vfs'] })
// one grep sweeps every source
await ws.execute('grep -rln session /redis /tmp')
// run a script that lives in Slack, report into Redis
await ws.execute('python3 /slack/report.py > /redis/report.txt')
// a typed CLI under a head word, found by man and which
ws.registerCli('slack', SLACK, { token: slackToken })
await ws.execute('slack send-message --channel general')

Virtual Terminal Overview

A Unified VFS
Virtual CLIs
The Mirage Shell
Virtual Runtimes
Policy Engine
Auth and Notifications
Profiles and Agents
Portability

A Real Example

A real Mirage workspace, running in your own browser tab. Mount GitHub and Redis, switch agent profiles, allow or refuse commands, hide folders and files, choose a Python runtime, install your own virtual CLI, and edit files in the terminal.

…

1mounts:
26runtimes:?The engines a line can run on. Both python engines are registered — monty and pyodide — and this entry says which one python and python3 bind to; quickjs takes node and js. All are sandboxed, so their file I/O only reaches the mounts above. The builtin command engine is always present, so it is not listed.
27  # Choose your own runtime for python, node, and anything else you run.
28  - { name:, captures: [python, python3] }
29  - { name: quickjs, captures: [node, js] }
30  # Route to e2b sandbox if python script has NumPy, see route.py
31  - name: e2b?Runs python that needs more than a browser has. Connect a store and a key, then put the script and its data where both sides can reach them:$ cp /repo/stats.py /repo/prices.csv /redis/$ python3 /redis/stats.py
32    captures: [python, python3]
33    config: { apiKey:  }
34 
35# Who takes the line, asked once before anything runs. It sits above the
36# bindings and below an explicit runtime argument.
37route_policy: ?A different door: asked once per line, before anything runs, and answering who runs it rather than whether it may. A runtime name places the line, a deny refuses it. It outranks the runtimes above and yields to a runtime named on the call.
38 
39# simple.js registers the CLI running in quickjs runtime.
40# On user alone: a cli is per workspace, and each profile has one.
41clis:
42  simple:
43    script: ?The program, run on the engine named beside it.
44runtime: quickjs
45 
46profiles:
47  admin:
48    commands:
49allow: [,,,,,,,,,,,,,, ]
50    policy:?A program the gate runs before every command this profile runs. It sees the command's context and reads the filesystem, and answers allow, ask or deny.
51      script: ?The program, run on the engine named beside it.
52runtime: monty
53 
54  user:
55    commands:
56allow: [,,,,,,,,,, ]
57    policy:?A program the gate runs before every command this profile runs. It sees the command's context and reads the filesystem, and answers allow, ask or deny.
58      script: ?The program, run on the engine named beside it.
59runtime: monty
60    mounts:
61      /tmp:
62        mode: read
63    paths:
64hide: [, ]

A real Mirage Virtual Terminal, running in browser?

…

A Unified Virtual Filesystem

A unified virtual filesystem over every source. S3, Google Drive, Slack, Gmail, Redis and the rest mount side by side under one root, so an AI agent can reach all of them through a unified interface with the Unix-like tools it already knows, like ls, grep, find and jq. A script sees the same root, and Python run through a virtual runtime opens those mounts exactly as the shell does.

VFS

…

Virtual Command Line Tools

A virtual command line for every service. git, slack and ntn are registered in the workspace rather than installed on the machine, so an agent drives each service with nothing installed, across different runtimes and machines. And one tool can be virtualized into two or more, each under its own name with its own credentials.

VFSVirtual CLI

…

The Shell

The shell is the glue. It binds the virtual CLIs and the Unix-like operations on the filesystem into one command line, so pipes, redirection, variables, jobs and history work across all three layers. A mounted Slack channel greps like a directory, its output redirects into Redis, and a virtual CLI reads the result back, in a single line an agent already knows how to write.

$ grep -rl VFS | Virtual CLI | Script

Virtual Runtimes

Runtimes are virtualized the same way, and not only for running scripts. Any command on the line can be redirected to a configured runtime, so Python might run in-process with Monty while another command, say kubectl, is sent to a remote machine over SSH. Which runtime serves a given line need not be fixed in the document either, since it can be chosen dynamically, per line, by a scripted runtime router.

$ grep -rl VFS | Virtual CLI | Script

Runtime

…

Policy Engine

A scriptable policy engine runs before every command in the virtual terminal. A policy decides which commands to allow, which to ask about, and which to refuse outright, and it decides where a command runs as well, routing it to a runtime from the context it can see. That context is the line itself, the paths it touches, and the virtual filesystem behind them.

Policy

$ grep -rl VFS | Virtual CLI | Script

Runtime

…

Authentication and Notifications

Mirage supports cloud authentication, and not only for what the virtual filesystem and the virtual CLIs need to reach a service. A credential store fills the virtual terminal's own environment variables too. It integrates with cloud notification services the same way, so an external change arrives as a change to a file and an agent reacts to it instead of rescanning the tree.

…

Authentication

…

Notifications

Policy

$ grep -rl VFS | Virtual CLI | Script

Runtime

Profiles and Agents

Mirage assigns a profile to an agent, and the profile is the whole of what that agent can see or do. It names the runtime, the mount mode for every path in the virtual filesystem, which command lines and runtimes the agent may reach, and which folders and files it cannot see at all.

Every profile is driven by an agent, and Mirage integrates deeply with the frameworks and harnesses teams already run, among them the OpenAI Agents SDK, the Claude Agent SDK, LangChain, Pydantic AI, CAMEL, Agno, OpenHands, Mastra, Haystack and more.

Authentication

Notifications

…

Profile 1…

Policy

$ grep -rl VFS | Virtual CLI | Script

Runtime

Portability

Mirage is extremely portable. It runs inside your own server, in FastAPI or Express, and just as well inside a browser tab, and it can become a machine's own filesystem through the mirage CLI or a real mountpoint over FUSE and FSKit.

Authentication

Notifications

Agents & Harness

Profile 1…

Policy

$ grep -rl VFS | Virtual CLI | Script

Runtime

…

Performance and Efficiency

Something interesting will come soon :)