Copy-paste patterns for real work — from your first run to headless CI to running a fleet of parallel agents. Every command here is accurate to the current mrc.
Point mrc at a repo. The first launch builds the image (~2 min); every launch after is ~5 seconds. Quit Claude and the container vanishes — your files never left the host.
The firewall already blocks the network path to your databases. .sandboxignore closes the other door — it makes the secret files themselves disappear inside the container.
Files are masked with /dev/null; directories get an anonymous overlay. Combine with the default firewall and there's simply nothing sensitive to find — on disk or on the wire.
Need web search and fetch? --web opens outbound HTTPS to any host — while the host-services lockdown and IPv6 block stay in place. Make it a default for a repo by dropping it in .mrcrc.
Need a specific private host but not the whole web? Add it to the allowlist loop in init-firewall.sh and rebuild — see the security docs.
Run a second session against the same project — say, one writing a feature while another investigates a bug. Each instance gets its own config volume and its own dynamically allocated proxy ports, so they never collide.
They share the same working tree — so mind edit conflicts. Two Claudes, one codebase, no good if they fight over the same file.
Same sandbox, different brain. Launch OpenAI's Codex CLI with --agent codex to cross-check a tricky refactor. It uses its own persistent config volume and the already-whitelisted api.openai.com.
Don't run one agent — command a team. Declare a roster (or grab a preset), mrc team up, and steer the crew from the dashboard or your phone. Each member is sandboxed with its own write territory; they propose, you steer, you commit.
They collaborate by @mention; a critic reviews risky work when called; you direct from the leads room and commit the work. Close a tab to suspend (reopen to resume); link a Telegram bot to answer from your phone. Full walkthrough → the teams guide.
Dynamic workflows can fan out many subagents at once. Give the VM room to breathe: mrc already defaults to all your cores and half your RAM, but you can push memory higher for heavy parallel runs.
Concurrency in dynamic workflows is capped at min(16, cores − 2), so once you're past ~18 cores more CPU won't add parallelism — but more RAM still helps when many agents run at once.
For CI or embedding, run detached with --daemon (prints a container ID) or stream machine-readable output with --json.
Name conversations as you go, then jump back into any of them. Reopening a repo resumes the most recent automatically; the picker lets you choose.
Commit a .mrcrc to a repo so everyone gets the same sandbox behavior. Personal preferences go in ~/.mrcrc; CLI flags override both.
Lost track of how many sessions are open across projects? mrc status sweeps every repo — and it needs no API key, so it's safe to alias and run anywhere.
"Ludicrous speed... GO!"— Dark Helmet
The recipes live in the repo. PRs with new ones are very welcome.