GitHub - brig-sh/brig: Brig: run coding agents in a microVM sandbox - the Go cli (brig) and session daemon (brigd)

Brig runs a coding agent inside a microVM on your own machine.

An agent working unattended can only damage what you handed it. Point it at one project, and a bad edit or a bad command reaches no further than that project. When you are done, throw the sandbox away and start clean.

How it works

One command starts a sandbox and runs the agent in it:

brig run claude ~/code/demo

A session is <agent> or <agent>@<label>, the ref every command takes. claude and claude@refactor are two independent sessions of the same agent, each with its own sandbox. The guest home is the host directory holding a session's settings and history. claude resolves to the claude-code agent, so its guest home is ~/.brig/homes/brig-claude-code, and claude@refactor's is the sibling ~/.brig/homes/brig-claude-code-refactor, not a directory inside it. Brig creates that home, and brig rm deletes it. Pass --home <dir> to use a guest home of your own, which Brig never deletes.

Name a project on the run line, and Brig mounts it read-write at /work/<name>, where the agent starts. Credentials reach the guest only when you deliver them: the sandbox boots with none, and the agent asks you to log in. What the guest does not get: every other host directory, your keychain, and your SSH agent. docs/sessions.md is the full model.

Requirements

macOS 15 is the floor: six of the eight built-in profiles need the hvi backend. See docs/install.md#platform-support for the rest.

Install

brew tap brig-sh/brig
brew trust brig-sh/brig
brew install --cask brig

The cask brings hull and cosign with it. For install.sh, Linux, or a source build, see docs/install.md.

First run

Each line is a check. A !! line prints its fix beside it.

mkdir -p ~/code/demo
brig run claude ~/code/demo

Brig prints brig: image and boot assets verified, Claude Code asks you to log in inside the sandbox, and pwd inside the agent prints /work/demo. The first run pulls the guest image and the boot assets, so it is slow. docs/authentication.md covers the login and how to carry one in from the host.

brig network publish claude 3000   # the agent's dev server, on localhost:3000
brig stop claude                   # stop the sandbox, keep its name
brig rm claude                     # stop it and remove it

brig rm also deletes the guest home Brig created. None of them touches ~/code/demo.

A published port binds to 127.0.0.1 and is named in the execution envelope, so the one inbound hole in the sandbox is one you asked for and can see. docs/quickstart.md walks through all of this, explained.

The boundary

The project mount is read-write, and those are your real files: the agent can change anything under it. On the default shared network the agent reaches the internet, so anything it can read it can also send. Brig enforces egress policy only on hull's hvi backend, and refuses a policy-bound run on any other backend rather than run it unenforced. Image verification defaults to warn, which reports an unverifiable image and boots it anyway. Set BRIG_VERIFY=require to refuse one instead. docs/security.md has the full picture.

Documentation

The full index is docs/README.md.

Project status

Brig is a prerelease, in the 0.1.0-rc series. brig version prints yours, and docs/stability.md says what you can script against today. docs/support.md says where to ask a question or file a bug, and SECURITY.md is where to report a vulnerability instead of a public issue. CONTRIBUTING.md covers the build, the tests and the review norms, and AI_POLICY.md says how AI-assisted contributions are handled. Brig ships under the Apache License 2.0. See LICENSE.

Brig itself counts nothing. On macOS the hull runtime it drives counts a few events, and on Linux nothing is sent. brig telemetry status shows the current setting, and brig telemetry off turns it off. docs/telemetry.md has the field-by-field detail.