Getting started

Connect a model provider

Fill in the API type, endpoint, API key, and model ID, see how Yenpo builds the request URL, and fix common connection errors.

The AI Reader uses whichever provider you’ve selected in Settings. Your provider account decides which models you can use and what you pay. Reading and searching files don’t need this connection.

Add a provider

Open Settings → Models, click Add Provider (the plus button), and select the new entry as the Current provider. Give it a name you’ll recognize, such as “Team gateway”. The name is just a label and doesn’t affect which API is used.

Then fill in:

Field What to enter
API type The protocol the service speaks: OpenAI Chat Completions, OpenAI Responses, or Anthropic Messages
Endpoint The service URL, including any path prefix your gateway requires
Authentication / API key Your API key, or the sign-in option offered for that provider
Model The exact model ID your account can use

Enter API keys only in Settings, never in a conversation or a project file. Some providers offer account sign-in instead of a key; if yours does, choose it under Sign-in method.

Match the endpoint to the API type

The model name doesn’t tell you the protocol: one gateway can serve similar-looking model names over different APIs. Follow the service’s API documentation and keep any required path prefix.

These URLs show how Yenpo builds the request URL. They’re examples, not working services:

API type Example Endpoint field Resulting request URL
Chat Completions https://gateway.example/v1 https://gateway.example/v1/chat/completions
Responses https://gateway.example https://gateway.example/v1/responses
Messages https://gateway.example https://gateway.example/v1/messages

If the endpoint already ends in /chat/completions, /responses, or /messages, Yenpo uses it as is for the matching API type. For Responses and Messages, an endpoint ending in /v1 gets the last segment added. Chat Completions always appends /chat/completions to whatever you entered, and it won’t add a missing /v1 for you.

If you get an error, don’t flip between API types until it goes away. Find out which protocol the service supports and what the full request path should be.

Model ID

Type the model ID into Model, or click Fetch List if the service publishes a model list. Fetching the list and calling the model are separate requests, so the model may work fine even when the fetch fails. Try entering the documented ID directly.

Context length affects Yenpo’s input and output budgets, when the conversation is compacted, and how much context is kept afterwards. Enter your model’s documented limit; leaving it blank uses Yenpo’s default. A value that’s too small can cause unnecessary compaction, while one that’s too large can lead to requests the model rejects.

While you’re getting the first connection to work, leave the other model settings alone.

Image input is off by default for each provider. Turn on Accept images only if the model and API accept the images you plan to send.

Test with a small question

Open a small file that you’re allowed to send to this service and ask:

What does this function return for an empty input? Cite the relevant branch.

If an answer comes back, the connection works. Whether the answer is right is a separate question, so check it against the code.

A request carries your message plus relevant code the AI Reader finds in the project, not just what you typed. Read Data, editing permission, and the terminal before sending confidential code to a remote endpoint.

When a request fails

Symptom What to check
Authentication or permission error The selected account or key, who owns the endpoint, model access
Not found or unsupported API API type and the full request path, including gateway prefixes
Model not found The exact model ID and whether your account can use it
Fetch Models fails Whether the service lists models at all; enter the documented ID instead
Timeout or interrupted stream Network access, proxy or gateway behavior, provider availability

An HTTP status alone rarely tells you the cause, and gateways may rewrite errors from upstream. When you report a problem, include the error message, the API type, and the endpoint with anything private removed. Never include the API key.

If a run with editing allowed was interrupted, check your working tree before retrying. Files may have changed even though no final answer arrived.