OpenComputer
OpenComputer is a cloud platform that gives AI agents a full Linux virtual machine instead of a short-lived container. Sandboxes persist for days, resize CPU and memory while running, and fork from named checkpoints.
What is OpenComputer?
OpenComputer is a cloud sandbox service built for AI agents, and its defining choice is architectural: every sandbox is a full Linux virtual machine with its own kernel, memory and disk, isolated at the hardware level through KVM rather than a container sharing the host kernel. The stated consequence is that untrusted, AI-generated code can be handed an entire computer without endangering the host.
The second choice is persistence. Most sandboxes are provisioned for a script that runs and dies; an OpenComputer machine keeps going for hours or days until it is explicitly killed, and its filesystem, installed packages and state survive between sessions. Idle machines hibernate and wake with everything where it was left. There is no session cap, which the vendor contrasts with rivals that stop at one or twenty-four hours.
Three capabilities sit around those ideas. Checkpoints are instant named snapshots that can be forked or restored in about a second, so five approaches can be tried in parallel from the same starting point. Elasticity means memory and CPU can be resized on a running machine, upward and downward, with no restart — the vendor's own comparison of eight providers presents this two-way resize as unique to it. Preview URLs expose a port publicly with authentication and custom domains.
The service is reached entirely through code: a TypeScript SDK, a Python SDK, an oc command-line tool and a documented REST API. A second product, Serverless Agents, sits on top: agents are written as reactive TypeScript functions with hooks for models, tools, subagents and MCP servers, then deployed to a managed cloud with durable sessions, immutable versions promoted through aliases, project secrets, managed egress, indexed logs and Slack channels. User-scoped Gmail, Calendar and GitHub connections are available.
The whole stack is open source on GitHub and documented for self-hosting, down to a GCP development cell. Twenty gigabytes of disk come with each machine and billing runs only while a machine is running. There are no GPUs, several features are still marked Preview, and the hosted service currently runs in a single unnamed region.
What it does
- Spin up a full Linux virtual machine in seconds from a TypeScript or Python SDK, the oc CLI or the REST API
- Run shell commands and read or write files inside the machine, with root access on a real x86_64 system
- Take named checkpoints and fork from any of them to try several approaches in parallel
- Resize memory and CPU while the machine is running, up or down, without a restart
- Hibernate an idle sandbox and wake it later in exactly the state it was left in
- Expose a port publicly through a preview URL, with authentication and a custom domain
- Define reactive agents in TypeScript and deploy them to the managed Serverless Agents cloud
When to use OpenComputer / When not to
A quick filter to help you decide if OpenComputer is the right fit.
When to use OpenComputer
- Teams building B2B agent products that need a computer their users' agents can keep, rather than a sandbox that dies after each task
- Backend and platform engineers running untrusted, AI-generated code who want isolation at the hypervisor level instead of a shared kernel
- AI and MLOps engineers whose workloads swing between idle and heavy, and who need memory and CPU resized on a running machine
- Developers automating long jobs — dependency installs, builds, test suites, multi-day agent runs — that no session timeout should interrupt
- Engineering teams with a hard self-hosting or bring-your-own-cloud requirement, since the whole stack is open source and documented for it
When not to use OpenComputer
- Anyone needing GPUs: the vendor's own comparison guide states plainly that OpenComputer has none and sends GPU work elsewhere
- Non-technical users: there is no no-code interface, no mobile app, and everything runs through an SDK, a CLI or the API
- Organisations with a strict data residency requirement, since the hosted service runs in a single region that is never named
- Buyers whose procurement needs terms of service, a data processing agreement, a subprocessor list or a SOC 2 or ISO 27001 certificate — none is published
- Workloads made of thousands of very short-lived environments, where start-up latency dominates cost and the vendor itself points to a faster competitor
How to use OpenComputer
A typical end-to-end flow, from setup to results.
- Install the SDK for your language: npm install @opencomputer/sdk, or pip install opencomputer-sdk
- Alternatively install the oc command-line tool, which drops into ~/.local/bin and needs no sudo
- Create an account on the dashboard and copy an API key from it
- Export the key as the OPENCOMPUTER_API_KEY environment variable
- Create your first machine with Sandbox.create(), which the quickstart claims takes about two minutes end to end
- Run commands inside it with sandbox.commands.run(), and read or write files with sandbox.files
- Take a named checkpoint once the environment is ready, so you can fork or restore from that point later
- Resize memory and CPU on the running machine when the workload demands it, then scale back down
- Expose a port through a preview URL if the sandbox needs to serve something publicly
- Kill the machine with sandbox.kill() when the work is done, since billing runs only while it is up
Pros & Cons
Pros
- Real virtual machine isolation with a dedicated kernel, the strongest tier in the vendor's own eight-provider comparison
- Two-way live resizing of a running machine, from 1 GB to 16 GB and back, with no restart or template rebuild
- No session timeout, where competing sandboxes cap sessions at one or twenty-four hours
- State persists natively between sessions, with no snapshot-and-restore pipeline to build and maintain
- Simple, published pricing billed only while a machine runs, with no monthly floor and a free tier to start
- The entire stack is open source and documented for self-hosting, which covers the bring-your-own-cloud case
- Substantial documentation: API reference, two SDKs, CLI reference, dated technical guides and an llms.txt index
Cons
- No terms of service anywhere on the site, so nothing contractually fixes liability, availability, termination or jurisdiction
- No legal notice, no full company name and no postal address, on either the product site or the publisher's own domain
- No data processing agreement, no named subprocessor list and no security certification such as SOC 2 or ISO 27001
- A single published contact, a personal address on a different domain, and no support email at all
- No GPUs, so GPU-bound work has to run somewhere else
- The hosted service runs in one unnamed region, which rules out any data residency commitment
- Two different price lists coexist on the site, and several advertised features are still marked Preview
Pricing & Plans
OpenComputer offers a permanent free tier, described as shared capacity of up to 2 vCPU and 2 GB of memory with 3 GB of disk and five concurrent sandboxes. Paid usage is metered by the running minute rather than sold as a subscription. The smallest machine, 1 GB of memory with 1 vCPU, is listed at USD 0.001 per minute, which the site's own calculator converts to USD 0.06 per hour and USD 42.18 per month of continuous operation; the default 4 GB machine is listed at USD 0.004 per minute. Twenty gigabytes of disk are included, with additional storage charged at USD 0.0000001 per GB-second. Volume discounts are offered on request.
- Free — shared capacity up to 2 vCPU and 2 GB of memory
- 3 GB of disk
- 5 concurrent sandboxes
- Usage
- 1 GB / 1 vCPU — USD 0.001 per minute
- USD 0.06 per hour
- USD 42.18 per month of continuous use
- Usage
- 4 GB / 1 vCPU (default machine) — USD 0.004 per minute
- USD 0.24 per hour
- USD 168.72 per month
- Usage
- 8 GB / 2 vCPU — USD 0.008 per minute
- USD 0.48 per hour
- USD 337.44 per month
- Usage
- 16 GB / 4 vCPU — USD 0.016 per minute
- USD 0.96 per hour
- USD 674.88 per month
- Medium — USD 300 per month for 8 CPU
- 16 GB of memory
- 50 GB of expandable disk and 50 concurrent sandboxes
- Large — USD 3
- 000 per month for 64 CPU
- 512 GB of memory
- 3
- 800 GB of disk and 500 concurrent sandboxes
- Custom — configuration and price on request
- Design Partnership — USD 100 per month with an 80% discount locked in on all future usage
Data, GDPR & hosting
A consolidated view of how OpenComputer handles your data.
GDPR overview
The privacy policy, last updated on 3 August 2026, mentions the GDPR but never claims compliance with it. It grants access, correction, export and deletion rights to everyone, then adds that depending on where you live you may hold additional rights under laws like the GDPR or CCPA, including objecting to processing and lodging a complaint with a supervisory authority. Rights are exercised by emailing a single address. Beyond that, the framework is absent: no legal basis for processing, no named controller, no data protection officer, no Article 27 representative in the European Union, no international transfer mechanism, no data processing agreement and no subprocessor list. There are no terms and conditions at all, and no postal address for the publisher anywhere. Silence on compliance, rather than a denial, is why this record marks GDPR compliance as not established.
Who owns the data?
The privacy policy is explicit on the point that matters most: the code, files and data run inside a sandbox belong to the customer. The vendor states it accesses that content only to operate the service, or with the customer's permission when helping to debug something. Account data — name, email address and details from the sign-in provider — plus usage metadata about sandboxes and agent sessions are held by the vendor to run the platform, meter billing and troubleshoot. Billing details pass through a payment processor, and full card numbers are never seen or stored. No terms and conditions exist to confirm, qualify or extend any of these commitments.
Reuse rights
Four uses are listed: operating the service, billing and fraud prevention, communicating about the account, incidents and product updates, and improving the product from aggregate usage patterns. Two exclusions are stated in plain terms — personal data is not sold, and the contents of sandboxes are not used to train machine learning models. Sharing is limited to the service providers needed to run the platform, each receiving only what its function requires, with disclosure to authorities where the law demands it. Non-essential emails can be opted out of at any time. Because sandbox content is presented as the customer's own data, nothing in the published documents restricts what a customer may do with it or requires permission to reuse it. The marketing site runs cookieless Plausible analytics and sees only page views, referrers and country-level location.
Data retention & training
Hosting summary
The privacy policy names no hosting country and no region. It refers to providers only by category — cloud infrastructure providers, a payment processor and email delivery services — stating that each receives only what it needs and is bound by its own data protection obligations. The vendor's own comparison guide says the hosted service currently runs in a single region, without naming it, so no data residency commitment can be read from the published material. On security, the policy states that sandboxes run in isolated virtual machines, that data is encrypted in transit, and that credentials and secrets are encrypted at rest, with a commitment to notify users of a breach without undue delay. An engineering post describes edge routing on Cloudflare Workers and per-second billing derived from ten-second heartbeats. Teams needing a specific jurisdiction have one documented route: the stack is open source and can be self-hosted on their own cloud account, with a GCP development cell covered in the documentation. No data processing agreement, subprocessor list or hosting certification is published.
Things to keep in mind
Risks and trade-offs to weigh before adopting OpenComputer.
- No terms of service are published anywhere, so there is no contractual footing for availability, liability, termination or the law that would apply to a dispute
- The publisher is identified only as Digger in a footer: no full legal name, no company number, no postal address, and the publisher's own domain has no legal pages either, which makes the counterparty hard to assess
- Billing runs by the running minute and sandboxes are designed never to shut down on their own, so a machine left up or an agent that forgets to kill one turns straight into an unnoticed bill
- Agents given a full machine with root access, stored secrets and outbound network can act far beyond what a human meant to authorise; the isolation protects the host, not the customer from their own agent
- No data processing agreement, no named subprocessor list, no certification and an unnamed hosting region make the product unsuitable for regulated or personal data until those documents exist
- A single personal email address on a different domain is the only published contact, which is a thin channel for an incident, a breach notice or a data request
- Several advertised capabilities are still marked Preview and the product is only a few months old publicly, so behaviour and pricing may change under workloads already built on them
Setup & Integrations
Technical difficulty
Setup is straightforward for a developer and out of reach for anyone else. Install the SDK from npm or pip, or drop the oc command-line tool into ~/.local/bin without sudo, then export an API key taken from the dashboard. A first sandbox takes three calls: create, run a command, kill. The quickstart claims two minutes. There is no graphical way to create a machine, and the dashboard mainly serves inspection and testing. Self-hosting the open-source stack, documented down to a GCP development cell, is a considerably heavier commitment.
Deployment
Integrations
Behind OpenComputer
Fundraising
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement OpenComputer.
Frequently asked questions
Is an OpenComputer sandbox a container?
How long can a sandbox stay alive?
Can I change memory and CPU while a sandbox is running?
What does it cost?
Is there an API?
Are GPUs available?
Is my data used to train AI models?
How long is my data kept?
Can I run it on my own infrastructure?
How do I contact the team?
Should you pick OpenComputer?
OpenComputer makes one clear technical bet and documents it well. Where most agent sandboxes are containers built for a script that runs and dies, this one hands each agent a real Linux virtual machine with its own kernel, keeps it alive for as long as needed, lets it be checkpointed and forked, and lets its memory and CPU be resized while it runs. The documentation backs the claims: an API reference, TypeScript and Python SDKs published to npm and PyPI, a CLI, dated technical guides, a public repository and a self-hosting path. Pricing is published, metered by the minute with no monthly floor, and a free tier exists to try it. For a team building agent products that need durable state and unpredictable resource demands, that combination is genuinely hard to find elsewhere.
The reservations are not about the product but about everything around it. There are no terms of service at all, no legal notice, no full company name and no postal address — the site names only Digger in its footer, and the publisher's own domain carries no legal pages either. There is no data processing agreement, no named subprocessor list and no security certification, and the hosted service runs in a single region that is never identified. Two different price lists coexist on the site without being reconciled. The product is also young: the domain was registered in February 2026 and the first published writing dates from March 2026, with several advertised capabilities still marked Preview.
The practical reading is that OpenComputer is convincing on engineering and thin on paperwork. Developers and technical teams can evaluate it on its merits today; anyone whose procurement, compliance or data residency requirements have to be met on paper should ask for those documents before committing.
- Choosing a selection results in a full page refresh.
- Opens in a new window.