Connect a self-hosted AI endpoint

A custom endpoint lets Boldo's assistant use a model service you provide. It changes the model provider inside Boldo; it is separate from the access granted to an external MCP client.

Administrators and the Owner can configure one custom endpoint per organization when the plan includes AI and custom endpoints.

Prepare the service

The service must be publicly reachable over HTTPS from Boldo. HTTP, localhost, private network addresses and redirects are rejected. A service accessible only on your computer or through your VPN cannot be used directly.

Obtain the base API URL, exact model or deployment name, required authentication and actual context-window, image and tool-call capabilities. Do not assume that declaring a capability in Boldo adds it to the model.

Configure and save

1. Open Organization settings > AI Assistant > Configuration.
2. Under AI configuration, click Add a self-hosted model.
3. Select Endpoint type: OpenAI-compatible or Azure OpenAI.
4. Fill the service fields described below, including the key when required.
5. Optionally enter a display name and open Custom headers > Add headers, then use Add header for each entry.
6. Enter Context window (tokens) and set Model can read images and Model supports tool calling to match the service.
7. Click Test and read the result. For a compatible service, Detected models (click to use) can help select the exact model name.
8. Click Connect to save. Testing alone does not save the endpoint.
9. Choose the custom model in Default model if you want it as the recommended default.
10. Open the assistant, select the model and check the service host shown in the selector. Verify a real response. If tools or images are enabled, verify an appropriate task as well.

Endpoint kind	Required service information
OpenAI-compatible	Endpoint URL, such as https://ai.example.com/v1, and Model name. A key is needed only if your service requires one.
Azure OpenAI	Resource URL, Deployment name and API key. Set API version (optional) to the version required by your deployment.

For a compatible endpoint, enter the base API URL, not the full /chat/completions URL. Boldo uses /models for its connection test and /chat/completions for conversation. For Azure, supply the resource URL and deployment name in their separate fields.

Azure requires a key, even if the generic key field appears optional. A model name conflicting with a reserved built-in model identifier is rejected. The context window must be a positive integer accepted by the form and must reflect the provider's real limit.

Interpret Test correctly

For an OpenAI-compatible service, Test checks the models endpoint. A successful HTTP response establishes reachability and authentication; it does not prove generation, streaming, image support or tool calling. An empty detected-model list does not necessarily mean the test failed.

For Azure, Test sends a minimal generation request to the configured deployment. The provider may charge for this request. This also checks the deployment, but does not certify every model capability.

An unverifiable result does not establish a valid key. Saving with Connect checks the configuration and destination, but does not substitute for a successful real conversation. If Test passes and chat fails, check the model name, paths and required capabilities.

Update stored credentials

Saved secrets are masked. When editing at the same address, leaving the key field blank keeps the stored key. For an OpenAI-compatible endpoint, use Remove key to remove it. Azure requires a key and does not offer this removal control. For an existing custom header, leaving the value blank keeps its secret; removing the row deletes that header.

If the host or URL path changes, re-enter the key and header secrets as requested. Boldo does not automatically forward saved credentials to a new endpoint address.

For a compatible endpoint, the dedicated key field supplies Bearer authentication and takes precedence over a custom Authorization header. Without a dedicated key, custom headers can supply the service's authentication. For Azure, the key field supplies api-key; a custom header does not replace the required key field.

Save the modified configuration, then test the chosen model again. Do not paste secrets into a support conversation or ticket.

Remove or troubleshoot the endpoint

Use Delete model and confirm to remove the Boldo configuration. This does not delete the remote service or conversation history. Review Default model afterwards; an unavailable choice falls back to Boldo recommendation. Stored credentials remain removable after a plan change removes the option to add them.

Problem	Next step
HTTP or private address refused	Provide an eligible public HTTPS endpoint.
Redirect refused	Use the final base API URL directly.
Provider rejects authentication	Check the dedicated key and authentication headers.
Key cannot be read	Re-enter the secret.
Test succeeds but chat fails	Verify exact model/deployment, API path, response compatibility and declared capabilities.
Model answers but cannot use Boldo tools	Check that it genuinely supports tool calls and that tool support is enabled.

A model without tools can handle ordinary conversation but cannot use Boldo's assistant tools; actions requiring them may refuse it. Custom-provider usage is not charged against Boldo's included hosted AI budget. Your service's own billing and limits still apply, along with Boldo's access and plan controls.