Moca
Moca is the rossoctl project for running AI coding agents safely and at scale. It has two capabilities. You can use either one on its own, or both together.
| Moca sandbox service | Moca harness | |
|---|---|---|
| In one line | Your coding agent stays on your laptop, and every command it runs goes to a fresh, isolated VM on a shared service. | Moca runs the agent itself: long-lived sessions on a warm worker pool, with state in Redis and tools in a remote sandbox. |
| Who uses it | Developers using Claude Code, OpenCode, Codex or IBM Bob | Teams and orchestrators that send agent work over HTTP, or users of mocactl |
| You drive it with | mocactl sbx | The HTTP API (/turn, /v1/runs) or mocactl |
| What you get back | The agent's edits, as a patch you apply (mocactl sbx pull --apply) | Replies, streamed or not, from sessions that resume by sessionId |
| Where it runs | You host it: an OpenShift cluster plus KVM hosts | A laptop (Docker Compose), a single VM, or Kubernetes / OpenShift |
| Start here | Sandbox quickstart | Harness quickstart |
See it step by step. Two interactive diagrams walk through these capabilities: the harness and the sandbox together and the sandbox service alone. Press Next to advance one step at a time.
How the two fit togetherβ
Both capabilities keep an agent's commands away from the agent's own process. They differ in where the agent runs.
- Shared parts. Both use moca's sandbox relay and its sandbox workers, which dial out to the relay, so a sandbox needs no inbound route. Both can run each command in its own Firecracker microVM. (Sandboxes and tiers, Install the sandbox service)
- One control plane, three profiles. The same control-plane component serves the harness, the
sandbox service, or both (
SH_ROUTE_PROFILE=harness,sandboxorcombined). Withcombined, one deployment serves sessions and sandboxes from a single template store. (One control plane for both) - Two different "sandboxes". A harness session runs its tools in the harness's own sandbox. A
sandbox-service workspace is a separate thing:
mocactl sbxcommands neither see nor manage a harness session's sandbox. The glossary defines both.
Which one do I want?β
| If you want to⦠| Use |
|---|---|
| Let a coding agent on your laptop edit and test code without running anything on your laptop | Sandbox service |
| Hand an agent a repository and get a reviewable patch back | Sandbox service |
| Run many agent sessions on a few warm processes, resumable after a restart | Harness |
| Dispatch units of agent work ("leaves") from your own orchestrator over HTTP | Harness |
| Run harness sessions and sandboxes on one control plane | Both, with the combined profile |
Not a fit today:
- A sandbox service on your laptop. The sandbox service deploys only to OpenShift plus a KVM host, so you, or your team, deploy it before anyone can use it. The harness has a five-minute local trial.
- Scheduled dispatch. The harness runs synchronous turns, streaming and detachable turns, human-gated runs and opt-in async fan-out. Scheduled dispatch has no design on the current runtime yet (#536).
- A hosted service. Moca is open source for you to self-host: there is no moca-run endpoint. Every deployment path is Preview, so interfaces and limits may still change (Maturity and lifecycle).
Find your pathβ
- Evaluating moca? Read the table above, then Get started, which says what each path needs. The harness quickstart runs on a laptop in about five minutes.
- Using the sandbox service? You need the URL of your sandbox service deployment. Start with the sandbox quickstart.
- Building on the harness? Start with the harness quickstart, then Use mocactl.
- Deploying or operating moca? Start with Deploy and operate. It says which deployment delivers which capability.
Statusβ
Moca is open source (Apache-2.0), for self-hosting, and in Preview. Where it is going, and the
evidence behind its design: Project. The latest release is
v0.10.0. Container images are
ghcr.io/rossoctl/moca and ghcr.io/rossoctl/moca-remote-worker. These pages track moca's main
branch.