OpenComputer

Product Introduction

1. Definition: OpenComputer is a managed cloud platform and infrastructure-as-a-service (IaaS) specifically engineered for deploying and running persistent, autonomous AI agents. At its core, it provides ephemeral, hardware-isolated Linux microVMs (KVM-based sandboxes) as the execution environment for each agent session.
2. Core Value Proposition: It exists to solve the infrastructure complexity of running stateful, long-lived AI agents that require a full, mutable operating system to execute tools (like git, ffmpeg, or browsers), maintain session durability, and operate securely with external APIs. Its primary value is being the "Firebase for agents" or the "background agent cloud," abstracting away VM management, session orchestration, and secure secret injection so developers can focus solely on agent logic.

Main Features

1. Agent-as-a-Function Runtime: Developers write an agent as a single TypeScript function. The platform manages the entire execution loop, including streaming, state persistence, and tool invocation. Deployment is via CLI (opencomputer deploy), abstracting all underlying infrastructure.
2. Full Linux MicroVM Sandboxes: Every agent session launches in a dedicated, hardware-level isolated KVM micro-virtual machine. This provides a real Linux environment with a read-write filesystem, shell access (bash, sh), package managers (apt, npm, pip), and full network egress, enabling agents to run any Linux-compatible tool or software.
3. Durable Sessions with Checkpoint/Fork: Sessions are not ephemeral containers; they can hibernate when idle and resume exactly where they left off. The underlying sandbox technology supports live checkpointing and forking, allowing for state capture and instant cloning of a machine's entire memory and disk state.
4. Secure Secret Management & Egress Binding: API keys and secrets are never exposed within the runtime VM. Secrets are bound to specific origin URLs (e.g., https://api.github.com) and are injected only into outbound HTTP requests from the sandbox, preventing the agent from reading or exfiltrating the credentials.
5. Bring-Your-Own-Model (BYOM) & Transparent Token Billing: The model (e.g., "anthropic/claude-sonnet-4.6") is specified as a configuration string. Users can bring their own API key or subscription, in which case OpenComputer passes model token costs through at the provider's API rates with no markup, billing only for machine time.
6. Bare Sandbox Access: For teams needing custom agent harnesses or runtimes, OpenComputer provides direct access to the same underlying sandbox infrastructure via a developer API (Sandbox.create()), offering full control over the microVM, including live resizing of CPU/memory, checkpointing, and forking.

Problems Solved

1. Pain Point: Traditional serverless functions (e.g., AWS Lambda) are stateless, short-lived, and lack a persistent OS environment, making them unsuitable for agents that need to maintain long-running context, install packages, or execute complex shell commands.
2. Pain Point: Managing self-hosted virtual machines or containers for agents involves significant DevOps overhead for provisioning, security patching, secret rotation, and cost optimization for intermittent workloads.
3. Target Audience: AI/ML Engineers and Agent Developers building sophisticated multi-step AI agents that interact with code repositories (via git), media (via ffmpeg), browsers (via Puppeteer), or other external tools requiring a full OS.
4. Target Audience: Product Teams integrating autonomous AI features (like automated code cleanup, content moderation, or data synthesis) into their applications, needing a reliable, secure, and managed backend for these processes.
5. Use Case: An automated code hygiene agent that clones a repo, scans for stale feature flags, and opens GitHub pull requests for cleanup, requiring git, filesystem access, and secure GitHub API token usage.
6. Use Case: A media processing agent that downloads, transcodes, and analyzes video files using ffmpeg and custom scripts, needing a persistent workspace and significant CPU bursts.

Unique Advantages

1. Differentiation vs. Container Platforms: Unlike generic container platforms (e.g., Kubernetes), OpenComputer is purpose-built for agents, with built-in primitives for durability (hibernate/resume), secure secret injection, and a managed agent loop, reducing boilerplate code.
2. Differentiation vs. Serverless AI Platforms: Compared to serverless AI platforms that focus on inference, OpenComputer provides a full, mutable Linux environment, enabling tool use that is impossible in pure, sandboxed serverless runtimes.
3. Key Innovation: The combination of lightweight KVM microVMs for hardware isolation with agent-specific session management and secret binding. This provides the security and flexibility of a full VM with the developer experience and cost model of a serverless platform, including per-second billing and scale-to-zero.
4. Key Innovation: The transparent, pass-through model token billing coupled with a compute-time-only meter when using your own key creates a unique and potentially lower-cost pricing model for high-volume agent workloads compared to platforms that bundle and markup model costs.

Frequently Asked Questions (FAQ)

1. How does OpenComputer's security model work for API keys? OpenComputer uses egress binding; secrets are stored centrally and injected only into outbound HTTP requests from the sandbox to a pre-defined origin URL. The agent runtime never has access to the raw secret value, preventing accidental leakage or exfiltration.
2. What is the difference between OpenComputer's "Agents" and "Sandboxes"? The "Agents" product is a fully managed runtime where you deploy a TypeScript function and OpenComputer handles the loop, sessions, and scheduling. "Sandboxes" are the raw, unmanaged Linux microVMs (the infrastructure layer) that power the Agents, available via API for teams who want to build and manage their own agent orchestration system.
3. How does the pricing work if I bring my own OpenAI or Anthropic API key? If you provide your own model API key or have a subscription (like Azure OpenAI Service), OpenComputer does not charge you for the token usage. You are billed only for the "machine time" (per-second compute of the sandbox). Your model costs are paid directly to the model provider.
4. Can I install custom software or system dependencies in an OpenComputer sandbox? Yes. Each sandbox is a full Linux microVM. You can use apt, npm, pip, curl | bash, or any other method to install packages, libraries, or even custom binaries during a session. The filesystem is writable for the duration of the session.
5. What happens to my agent's state when it hibernates? The entire state of the microVM (memory and disk) is checkpointed to persistent storage. When the session is resumed, it is restored exactly from that point, with all variables, file handles, and process states intact, as if it had never stopped.