Open-source workbench for AI coding · macOS · Linux · Windows

Approve every command.
Let the agents run.

Give your AI tools their own workbench: a real container with your whole stack, services and repos, from one file. Turn them loose, and whatever they do stays off your machine, where your client data lives.

Open source · your own Docker · no cloud, no account, no rental

Today

AI on your real machine.

  • Babysit every step, or leave it running and hope
  • Auto-accept could reach your keychain, your tokens, your other clients' code
  • One bad command takes the whole machine with it
  • Cleaning up means hunting down everything it touched
With Monoceros

AI in its own workbench.

  • Let the agent run unattended, start to finish
  • Auto-accept freely: your machine, keychain and other clients stay out of reach
  • Breaks stay in the workbench, not on your machine
  • Your whole stack in one file, reset to clean in one command
Getting started View on GitHub
Several worlds, one machine

Three clients. Three workbenches. One laptop.

Each project gets its own container: its own stack, services and agent. What happens in one never leaks into another, or onto your machine.

client-a
node · postgres · claude
client-b
java · mysql · rovo dev
side-project
python · redis · claude
Workflow

Four commands, start to finish.

Describe the workbench, build it, let the agent run, reset to clean. That's the whole loop.

01

Describe

Say what goes in: your stack, your repo, the AI tool. One YAML, editable.

 monoceros init acme \
    --with-languages=node,java \
    --with-services=postgres \
    --with-features=claude,github,atlassian/twg \
    --with-repos=https://github.com/acme/booking-manager.git
02

Build

A real container with your stack, up in minutes. That's the box your agent lives in.

 monoceros apply acme

  ▸ pulling image, installing features...
  ✓ workbench ready
03

Let it run

Hand the agent a task and let it run, unattended. It works the repo inside the workbench - and can't reach past it to your machine.

 monoceros run acme --in projects/booking-manager -- claude "Implement Story ACME-345"

  ▸ agent working in acme...
04

Reset

Made a mess? Reset to clean in minutes. Or keep the workbench for next time.

 monoceros remove acme

  ▸ snapshot saved
  ✓ host clean
Comparison

Other ways to run an agent today.
And the catch with each.

The common ways people run AI agents today, and the catch with each.

The agent on your real machine
Full access, full risk.

It can touch your keychain, your other clients, the whole disk. So you babysit every step - or flip on auto-accept and hope.

Monoceros
Contained from the start.

The agent runs in its own workbench. Turn it loose - it can't reach the rest of your machine.

Cloud agent platforms
Isolated, but not yours.

Your code and client data live on someone else's servers. Off-limits for many contracts, useless offline.

Monoceros
The same isolation, local.

Runs on your machine, not a vendor's. No cloud account, no rental - and it works offline.

Roll your own container
Works, but it's all on you.

A container or VM per project, wired up by hand. No lifecycle, no clean reset - and you maintain it forever.

Monoceros
The box is the product.

One YAML, a real workbench, reset in one command. You don't build or maintain the box - that's our job.

Local agent sandboxes
A box for the agent, and a bill.

The newer agent sandboxes isolate well. But they are proprietary and paid, want an account, and give you just a box: your stack, services and repos are still yours to wire up.

Monoceros
The whole bench, open and free.

Open source, on your own Docker, no account. One file builds the whole environment - languages, services, repos and your agents - not just a sandbox to drop one into.

What you get

Your stack. Your tools. Your machine.

Pick what goes in the workbench. Monoceros wires it up - reproducible, isolated, local.

Your languages

Node, Python, Java, Go, Rust, .NET. Pin the version per workbench.

Your services

Postgres, MongoDB, Keycloak and many more. Each workbench gets its own, isolated.

Your tools

Claude Code, Rovo Dev, GitHub CLI, and many more. Each one a feature you add - swap freely, no lock-in.

Your repos

Point at an HTTPS repo - GitHub, GitLab or Bitbucket. One token clones it and signs in gh / glab inside the container. No git tooling on your host.

Reproducible, pinned

One YAML builds it, the same on any machine. Each workbench pins its runtime version; apply never changes it behind your back, upgrades are a deliberate command.

Local, no cloud

Your Docker, no account, no rental, no telemetry. Runs on your machine, offline included.

Open standard underneath

Built on the dev-container spec VS Code and others use. Walk away anytime.

Open in your IDE

Attach VS Code, VS Codium or JetBrains to the running workbench over SSH - zero setup, Windows included. One monoceros open and you are editing inside the container.

Logins that stick

Tool logins, sessions and editor extensions survive a rebuild. Run apply again, and your agents stay signed in, your setup intact.

See all features → More tools, languages and services coming soon.

Get started

Run your first agent.

Open source, runs on your machine. Install on macOS, Linux or Windows, then one command to build a workbench.

macOS / Linux
curl -fsSL https://raw.githubusercontent.com/getmonoceros/workbench/main/installer/install.sh | bash
Windows (PowerShell)
irm https://raw.githubusercontent.com/getmonoceros/workbench/main/installer/install.ps1 | iex