Connect tools and integrations
A model can produce a text response without being able to read a workspace file, run a command or contact another service. For those operations, the relevant tool, connection, execution environment and permissions must be available.
On this page
The interface steps below describe the October 2026 development version. The v0.0.1 release distributes the Server separately from runtime/plugin assets; installing the Server alone does not establish tool readiness.
Find the tool you need
- Open My tools for the tools visible to your account. Administrators also have Tool admin for their management scope.
- Search for the operation or tool name and open its details.
- Check whether it is available or Not ready, and read any readiness reason.
- Inspect the tool's input description, source, declared output and permissions and approval rules. Note any allowed hosts or approval requirements.
- Decide whether the task actually requires it. Prefer a text-only response when that meets the goal.
A tool listed in the catalog is not necessarily ready to execute. A model connection test does not test that tool's external service, credentials or runtime dependencies.
Load a tool for a session
Where session management is available, open Advanced settings → Runtime config and use Loaded tools to select the intended available tools. Confirm the selection and Save.
Selection is not a new permission grant and does not install missing software. It describes the tools loaded into that session; do not treat this selector alone as a complete security allowlist for every possible operation. Keep the task's allowed resources and actions explicit, and use the instance's access and approval controls.
Check the environment before a real task
Start with a small operation whose effect you understand. For a file-reading tool, ask it to inspect a non-sensitive sample you intentionally supplied and return an answer in the conversation. State that it must not change files or contact another service. This is a suggested verification task, not a guarantee that every tool supports read-only operation.
Inspect the tool result as well as the assistant's explanation. A displayed tool call is not proof of success, and a successful call is not proof that the final answer meets your request.
For an external connector, confirm the connected account and intended resource before authorizing a write or send. Never use production changes as a connectivity test.
Understand Runtime Assets
Runtime Assets provide software used by execution and processing capabilities. They are not the chat model and are not interchangeable with a skill. An administrator may need to configure the execution environment or install compatible packages/plugins for the required operation.
The management interface provides a Download source control for the Runtime Assets catalog. A customized catalog changes where packages are obtained. Only an administrator who has verified that source should change it; copying an arbitrary catalog URL from a forum is not a repair procedure. This guide does not prescribe an unverified installation command for missing assets.
Diagnose the layer that failed
- No matching tool: check whether it is available to your account and whether the required connection/plugin is configured.
- Not ready: read the stated reason; resolve the named dependency or connection rather than changing unrelated model settings.
- Authentication or permission error: verify the intended account and resource. Do not broaden access simply to make the error disappear.
- Approval requested: inspect the actual operation and scope before deciding.
- Tool returned an error: keep a redacted error and identify whether anything completed before retrying.
- Tool succeeded but answer is wrong: compare the result with the source and task requirements; this is not necessarily an installation failure.
For support, include the tool name, version, operation, readiness message and minimal redacted error. Do not send secrets, raw authentication headers or a full unreviewed log dump.
Use model configuration for connection problems and human oversight for approvals and handover.