Assistant

The Assistant is a bring-your-own-key chat rail on the right side of the Desktop window. It answers questions about your local AgentKit usage from a short aggregate summary, using an API key you store for Anthropic, OpenAI, or OpenRouter. After this page you will know how to set a provider and key, what a message sends, and what the errors and limits mean.

Open the rail

Select Open AI Assistant at the right edge of the window to expand the rail, and Collapse AI Assistant to close it. The rail appears only when the window is wide enough, and Desktop remembers whether you left it open.

Choose a provider and model

Provider starts on Anthropic. Model is a free-text field for every provider: for Anthropic it suggests claude-sonnet-5 and claude-opus-5-5, and for OpenAI and OpenRouter you type the model id yourself. A model change is saved when you leave the field. Switching providers resets Model to that provider's default.

Desktop saves the provider and model, never the key, in assistant/settings.json under your AgentKit home.

Store your API key

Paste the key into API key and select Save (or press Enter). The field is write-only: it clears as soon as you save, and the key is never shown again. The rail only tells you A key is stored for this provider. or No key stored yet. Each provider keeps its own key. Pasting into Replace stored key overwrites it, and Remove key deletes it.

The key goes into the local secrets vault as assistant-<provider>-api-key (for example assistant-anthropic-api-key), so it is listed on the Secrets page like any other stored name.

Check what leaves your machine

Select What leaves your machine to see the exact summary the next message sends as context, plus the text you type, to the selected provider. The summary holds only:

• the number of sessions in the last 7 days;
• up to five most-used skills in the last 7 days, by name;
• the kinds of recent activity events, such as run.started; and
• whether the local analytics index is enabled, and its health.

No file paths, transcripts, or secret values are included. Nothing is sent until a key is stored and you press Send.

Chat

Type in the message box and select Send, or press Enter (Shift+Enter adds a new line). The reply streams into the rail as it arrives; Stop ends a reply in progress without showing an error. Each message sends the whole conversation so far, which Desktop keeps only in the open rail. New chat clears it, and collapsing the rail also ends it.

To send a message, Desktop reads the key from the vault. That read is recorded in the secrets audit log with the reason assistant: send chat message, and Desktop reuses the key in memory for up to 15 minutes instead of recording a read for every message. Changing the key or provider, or locking the vault, drops that copy.

Errors and limits

Provider failures appear as one plain sentence in the rail:

One message waits at most 60 seconds for the provider to start answering and 2 minutes in total. A reply is capped at 2,048 tokens. A conversation may hold up to 500 messages, each up to 32 KiB. Provider calls honor the HTTPS_PROXY and HTTP_PROXY environment variables.