# BaseDock

> BaseDock is a separate, governed place for your team to build software with AI agents, with real apps, databases and VMs to work on. It has its own login, network and bill, separate from the systems you run the business on, and every action an agent takes is recorded and risk scored.

BaseDock is in beta. Join at https://portal.basedock.uk.

## Why it exists

An agent that can only write code is a demo. To be useful it has to deploy the thing, stand up a database, break it and try again. That needs real infrastructure, and the obvious source is the cloud you already pay for.

Azure, AWS and GCP all offer dev and test environments, and with the right governance layered on top they are well protected. That governance is something your team designs, builds and maintains alongside everything else it runs.

BaseDock starts from isolation instead. It is a separate platform with the governance already built in, so an agent can start work minutes after you hand it a key, and what it builds leaves with you as ordinary containers.

## Where agents can build

| | A dev environment in your cloud | A hosted agent platform | BaseDock |
|---|---|---|---|
| Separation from daily operations | Separated inside your account by the governance you add | Separate, on the vendor's infrastructure | A separate platform, isolated by default |
| Governance | Added by your team on top | Provided by the vendor | Built in: roles, resource quotas, firewall rules and written instructions your agents are handed |
| What caps the bill | Budgets and alerts you configure | Plan and usage limits set by the vendor | Hard quotas on resources and network traffic, set before anything runs |
| What you can take with you | Built on the provider's services | Your code, with the environment on the vendor's platform | Ordinary containers and standard databases |

## What it changes for engineering, product and infosec

### Engineering

Use agents properly, without waiting for a platform team to build somewhere to run them.

- Real compute: container apps, VMs, PostgreSQL, MySQL, Redis and MongoDB, each group on its own network.
- Claude Code, Cursor, Windsurf, VS Code and Devin connect directly, through MCP (the Model Context Protocol, the standard way agents plug into tools), and can create and run resources themselves.
- A working method agents pick up on arrival: research, challenge, then implement, with each change broken into pieces you can review.
- Opinions already made. Naming, quotas, isolation and access come configured, so the first week isn't spent deciding them.
- Standard container images run here unchanged, and they run somewhere else later. Images that need direct kernel features, such as FUSE or nested Docker, don't fit the sandboxed runtime.

### Product

See ideas running, and know what the agents did to get there.

- An idea becomes a running HTTPS address you can put in front of users, exposed in one click on the managed edge.
- A finished piece of work ends in a plain summary of what changed, what was assumed, the risks and what comes next.
- Write the team's rules once, such as coding standards, review steps and how agents should behave, and apply them to a project, a team or the whole company. Every agent connected through MCP is handed those rules, and a project can require agents to record the working method before they deploy.
- Hard quotas mean an experiment cannot grow past the limits you set for it.

### Infosec

Keep agent work contained and visible, without opening another review queue.

- Nothing to mix up. BaseDock has its own login, network and bill, separate from your production systems.
- Each project runs on its own network and outbound traffic is filtered. Every workload runs on a sandboxed runtime (gVisor), and standard containers are refused for tenant code.
- Every change by a person or an agent lands in an append-only audit trail with a risk score. A read-only auditor role sees it without being able to touch anything.
- Keys are limited to what they need, can be revoked at any moment, and are shown only once when created.
- Governance is there from the first login, so there is no retrofit project later.

## What we decided for you

- **Every agent gets only the access you give it.** Each agent works with its own key. You can limit it to one project or to deploying only, and revoke it at any moment. It can never do more than the person who issued it, secret values it sets are masked, and everything it does is recorded under its own name.
- **Containers, not a proprietary runtime.** What you build runs anywhere containers run, so you can take it with you.
- **Visibility before control.** Agents get room to work. You get to see all of it, as it happens.

## What the audit trail looks like

Illustrative entries. Risk runs from 0 to 100 and is banded info, low, medium, high and critical. The real trail also records the resource and the client IP.

```
14:02:11  claude-code   db.create       postgres/ledger-test           low 18       ok
14:02:40  claude-code   app.deploy      shop-api:0.4.2                 low 24       ok
14:03:05  claude-code   ingress.create  0.0.0.0/0 port 5432            critical 90  denied
14:03:06  alice@acme    quota.raise     cpu 4 to 8                     medium 35    ok
```

## More

- [Security and governance](/security.md)
- [Questions and answers](/faq.md)
- [Terms of use](/terms.md)
- [Privacy policy](/privacy.md)
