mirage

The World's First Virtual Terminal for AI Agents

What is mirage?

The World's First Virtual Terminal for AI Agents

Overview of mirage

What it is

Mirage is a virtual terminal for AI agents. It provides a single shell interface over data, command-line tools, and runtimes. Unix tools like grep can reach multiple services, each using its own native search. Commands are routed to the most appropriate runtime, with most lines executing in-process and only heavy work sent to a sandbox, reducing infrastructure cost and token usage. A workspace consumes about 100 KB of memory, and each new session about 10 KB, allowing one process to serve many agents simultaneously. Profiles control what an agent can run and see, and scriptable policies enforce rules.

Key features

  • Unified virtual terminal interface: every backend uses the same filesystem semantics, enabling pipelines across services.
  • Virtual filesystem over multiple sources: S3, Google Drive, Slack, Gmail, Redis, and others mount side by side under one root, accessible via standard Unix tools.
  • Virtual command-line tools (CLIs): tools like git, slack, and ntn are handled by Mirage itself, requiring no installation on the agent.
  • Routed, dynamic runtimes: Python, JavaScript, and other commands can be sent to in-process, sandboxed, or remote runtimes, decoupling computation from storage.
  • Lightweight design: a workspace takes about 100 KB of memory, each new session about 10 KB.
  • Profiles for agents: allow, ask, and deny govern commands and CLIs; hide and show govern files and folders.
  • Scriptable policy engine: policies can prohibit dangerous actions before execution, and the same stack gates every VFS operation and session write.
  • Notifications wired into the VFS: external changes become an event stream on the mount, so agents react to changes without rescanning.

Use cases

  • AI agents that need to search and manipulate data across multiple services (e.g., S3, Slack, Gmail) using a single command-line interface.
  • Running scripts that live in one service (e.g., Slack) and output results to another (e.g., S3).
  • Applying fine-grained access control to what an agent can see and execute, with hidden paths and denied commands.
  • Routing computation to different runtimes (in-process, sandboxed, remote) based on the command or library used.

Getting started

See the repository README for installation and quick-start instructions.

License and pricing

Licensed under Apache-2.0. Open source.

Based on the project's public README and repository metadata.

Add this badge to your README

Featured on YunoDB Badge styles & help

Maintain this project? Show that it is featured on YunoDB.

Markdown[![Featured on YunoDB](https://yunodb.com/badge/strukto-aimirage.svg)](https://yunodb.com/tools/strukto-aimirage?utm_source=github&utm_medium=badge)
HTML<a href="https://yunodb.com/tools/strukto-aimirage?utm_source=github&utm_medium=badge"><img src="https://yunodb.com/badge/strukto-aimirage.svg" alt="Featured on YunoDB" height="20"></a>