Integrations
MCP & connections
Understand planned connection surfaces and integrate against the existing gate today.
The intended MCP flow
A compatible AI tool would request an operation through Kernall, which would check scope and policy, pause when review is required, and return a result from the execution environment.
No MCP tool names, transport settings, OAuth scopes, or setup snippets are released. Adding an arbitrary endpoint to an MCP client does not create a server.
Integrate your own agent today
Wrap a tool call with the HTTP API or Python client.
- Find where your application performs an external operation.
- Map it to an action type, destination, and relevant data.
- Make a check with a stable step ID.
- Stop, wait, or execute according to the decision.
- Record the operation’s actual outcome in your application.
This is an adapter you implement, not a supplied MCP gateway. It does not prevent access through other tools by itself.
GitHub, Drive, cloud, and models
| Connection | Current status |
|---|---|
| GitHub | Authorization, repository import, and reviewed branch/PR export are planned. |
| Google Drive | Authorization, selected-file import, and reviewed export are planned. |
| AWS / Azure / notebooks | Runtime adapters and compute lifecycle controls are planned. |
| Kernall-hosted models | No model gateway, catalog, or inference endpoint is implemented. |
| Teams | Shared-project roles and teammate approval permissions are planned. |
What connection setup will cover
A released integration needs an exact setup path, requested permissions, supported operations, credential handling, failure recovery, and disconnect behavior. Cloud integrations also need usage and shutdown controls. These are release requirements, not current guarantees.