- Developers keep the terminal and IDE they use today and point an open coding agent, OpenCode, at an Enterprise h2oGPTe endpoint inside your network.
- The agent helps with the daily work: writing and refactoring code, reviewing pull requests, writing tests, drafting documentation, and explaining unfamiliar code.
- It answers from your own repositories, internal API docs, design records, and runbooks, and it cites what it used.
- H2O Cascade fine-tunes open code models on your repositories, so the model learns your conventions without the code going to an outside service.
Your developers want AI help, and your code cannot leave
Your engineering teams see what AI coding tools can do and want them in their daily work. Your security team will not let source code, internal API specifications, or design records go to an outside coding service. In some networks that service is not reachable at all.
So you end up with two groups. Some developers use a cloud assistant under a policy exception, with per-seat spend that keeps growing. Others, often on the most sensitive code, get no AI help at all.
The risk is also in how agents are run. A coding agent that runs code needs identity, access control, and an audit trail from the first day.
How a sovereign AI coding assistant works
Enterprise h2oGPTe serves models through governed endpoints inside your environment. OpenCode, an open-source coding agent, runs in the terminal and in IDEs such as VS Code and JetBrains, and becomes an agent harness you run yourself when pointed at your Enterprise h2oGPTe endpoint.
Step 1: connect the agent to the gateway
Developers point OpenCode at the Enterprise h2oGPTe gateway, never at an internal model endpoint. The gateway handles sign-in through your identity provider, role-based access, and audit.
Step 2: run tool calls in isolation
When the agent reads files, runs tests, or executes commands against a source tree, it does so in isolated agent execution, apart from other users.
Step 3: index what your code depends on
Repositories, internal API docs, architecture decision records, and runbooks go into private collections. Hybrid lexical and neural search finds exact function names, config keys, and error codes.
Step 4: answer with sources
When the agent explains a module or proposes a change, it cites the files and documents it used, so the developer can check.
Step 5: review and merge as usual
The agent proposes; a developer reviews every change in your normal code review before it merges.
Step 6: tune on your own code
H2O Cascade adapts open code models to your repositories and conventions, so the model learns your patterns without the code going to an outside service.
What developers use it for day to day
The agent helps write and refactor code, review pull requests, write unit and integration tests, and draft documentation, and it can explain unfamiliar parts of a codebase. Some code, such as core banking systems, network configuration, clinical integrations, and mission systems, is off limits to every outside service, including on air-gapped networks. This is the work a sovereign AI coding assistant is built for.