Changelog

Follow new updates and improvements to BoltAI.

Build 304

BoltAI 3.1 keeps long chats going with Auto-compact and reads scanned PDFs.

New

Auto-compact. When a chat gets long, BoltAI summarizes the earlier messages so you can keep going. New chats use it unless you choose another Context Policy. [ref]
Type /compact to summarize a chat right away. [ref]
A ring next to the mic shows how much context your next message sends. Click it for details. [ref]
Scanned PDFs. When you attach one, BoltAI offers to read it with OCR.

PDFs with no readable text show a “No text” warning in attachment lists. Click it to read the PDF with OCR.
Token counts are closer for code, indented text, and Chinese, Japanese and Korean text.

Build 97This update improves how chat history and attachments are sent to the model.

Improved how chat history is sent to the model in long chats.
Chat Context now defaults to the last 20 message pairs instead of 8. Chats, projects and agents with their own setting keep it.
In the Inspector, settings that come from a project, an agent or your defaults are now shown read-only. Click Edit… to customize the chat or change the settings where they come from.
Long chats that send all messages now show a notice above the composer, with an option to send only the last 20 pairs.
Large images are scaled down to fit provider limits when sent, in their original format. The attached file isn't changed.

Fixed Chat Context applying only to the first request when the model uses tools. Later steps sent the full chat.
Fixed attachments the model can't read being dropped silently or sent as files the provider rejects. The model is now told which file it couldn't read.
Fixed an image that couldn't be loaded being left out of your message. The send now stops and tells you.
Fixed spreadsheet attachments reaching the model without the values Excel shows, such as dates and percentages.
Fixed RTF attachments being sent as RTF markup instead of text.
Fixed Quick Compose and the Instant Chat Bar dropping your default system prompt and parameters when you picked a model.
Fixed agents set to "Use Default" Chat Context ignoring your account's Chat Context.
Fixed a reply you stopped being sent back to the model with its unfinished reasoning. Replies cut off by the output limit keep their reasoning.
Fixed Save & Regenerate on an earlier message regenerating the wrong reply while the chat showed an error.

Build 96This update fixes images and sign-in for MCP connectors, Anthropic chats that use web search, and chat titles on on-device models.

If you forget your master password, Settings › Security now has a "Forgot Master Password?" option to reset it, and the steps in the password prompt are correct.

Fixed images returned by MCP tools reaching the model empty. Images now show on the reply, and Anthropic and Google models can see them. Other providers get a note that an image came back.
Fixed a large MCP tool result leaving the chat stuck on "Running tool...", and large images slowing the chat down while the reply streams.
Fixed MCP connectors that also work without signing in, such as Seats.aero, never showing their sign-in page. Connect and Reset Authentication now open it. [ref]
Fixed some Anthropic chats failing on every message with a "tool_use ids were found without tool_result blocks" error after Claude used web search. [ref]
Fixed chats on Apple Foundation Models, Ollama and LM Studio sometimes getting the first words of the message as their title. [ref]