OpenComputer

OpenComputer, from the team behind Digger, is a background-agent cloud: a managed runtime where an agent is an exported TypeScript function rather than a server you have to operate. Hooks such as useModel() and useInput() attach capabilities, the function's return value becomes the instructions for a managed agent loop, and OpenComputer takes care of the parts that usually eat the sprint — the loop itself, durable sessions, streaming, tool wiring and production versions.

The runtime

There is no local agent server. Code in a project's opencomputer/ directory is watched and synced into OpenComputer's managed development cloud, so the environment you test against is the environment you ship to. Immutable deployments are promoted through aliases such as production, and sessions can be replayed and inspected long after they finish.

Underneath, every sandbox is a full Linux microVM with hardware-level isolation via KVM — shell, filesystem, package manager and network, not a language-level sandbox. Sessions are durable: they stream, can be steered mid-run, hibernate when idle and resume where they left off, and machines can be checkpointed, forked, and resized in memory and CPU at runtime. Teams that want the primitives without the framework can drive bare sandboxes directly and own the loop themselves.

Secrets that never enter the agent

The security model is the part worth reading twice: project-level secrets, with optional per-agent overrides, are bound to a single origin and injected after the request has left the sandbox. Agent code can call a secret-backed API without ever holding the credential, which removes the usual prompt-injection route to key exfiltration.

Getting started

The SDK is @opencomputer/agent with @opencomputer/cli for project management; agents can call tools, MCP servers, skills and subagents. A dashboard and CLI playground let you drive sessions interactively before wiring anything up.

Pricing

Pay-as-you-go with $10 of free credit: model time at API rates and machine time billed per second, from $0.00315/min for the default 2 GB / 1 vCPU baseline up to $0.05040/min for the largest configuration. Pro is $20/month for 10x prepaid credit, Max is $200/month for the same deal at scale, and enterprise terms are custom.