The Terminal Environment for AI Agents, Everywhere

Agents already know the terminal. We can give them a virtual one.

Mirage

The World's First Virtual Terminal for AI Agents

Mirage brings a unified virtual filesystem, virtual CLIs, virtual machines and policies. Each virtual terminal starts with only KBs of memory.

adminuser
$

Cloud search, built into the Unix tools an agent already uses.

What the Mirage animation shows, chapter by chapter.

  1. Virtual Filesystem. Cloud search, built into the Unix tools an agent already uses.

    • grep -rlw Northwind /gmail /gdrive /slack Plain grep -r, answered by each app's own search index, reads 3 files instead of 6,412.
  2. Virtual CLIs. Services as command lines Mirage answers itself, with nothing to install.

    • gh pr list --label bug Mirage answers gh itself, so nothing is installed.
    • wiki read-wiki-structure '{"repoName": "strukto-ai/mirage"}' | head -4 An MCP server becomes a virtual CLI, with each tool as a subcommand.
  3. Policies. Two agents share one workspace. Each profile decides what its agent sees and runs.

    • ls -A /jobs admin's profile has no rules, so it sees the secrets beside the code.
    • ls -A /jobs user's profile hides *.env and *.pem, so to user the secrets aren't there at all.
    • rm /jobs/clean.py user's allow list has no rm, so the line never runs.
    • gws gmail send --to lee@rival.com --subject "Q4 pricing" Policy holds mail leaving acme.com for a person, who rejects it.
  4. Virtual Machines. One shell. Each python line runs where its imports need.

    • python3 -c "print(2 ** 10)" Plain Python stays in this process, on Pydantic Monty, with no machine to start.
    • python3 /jobs/clean.py This script imports numpy, so the routing policy sends it to a Microsandbox microVM.
    • python3 /jobs/finetune.py This one imports torch, so it goes to an AWS EC2 GPU instance.
  5. Shell. Jobs, caches and logs, kept on Redis and disk.

    • python3 /jobs/finetune.py & & starts the training as a background job, and the store keeps it in Redis.
    • jobs Its output streams into that Redis, so any other process can follow it live.
    • jobs The stream ends with the job's exit code.
    • history 4 Every line is a command log on disk, and a trace in an OpenTelemetry backend once the process sets up its SDK.
AgentVFS

The Virtual Filesystem Interface for AI Agents

An open spec, VFS.md, and its benchmark, VFS-Bench 1.0. A fraction of MCP's tokens, and more tasks done.

MCPRequests 0Tokens 0.00M
VFSRequests 0Tokens 0.00M
$
Token reduction
Up to 88%
Average success rate increase
+5.5%
  • Eigent AI
  • CAMEL-AI
  • Zilliz
  • Milvus
  • TiDB
  • MongoDB
  • Redis
  • microsandbox
  • MIT
  • Harvard University
  • University of Oxford
  • Yale University
  • UIUC
  • Columbia University

Team From

AWS AIStanford UniversityYCombinator