Bolt Forge

Bolt Forge is a third agent inside Bolt.new, alongside the existing Standard and Max agents, and it runs only on open-source models: GLM 5.3 Flash, which the company recommends for most work, GLM 5.3, and Kimi K3 and DeepSeek v4 Pro as experimental options that consume an allowance faster. It was announced on 14 September 2026 as a research preview running to 14 October, and reached a wider audience through a product launch post five days later. The trade at the centre of it is stated plainly by the company, which is what makes it worth listing: open models cost less to serve, so Forge offers far more usage — up to fifty times more for a Pro subscriber at no extra cost during the preview — and in return the user must consent to their session being used as training data. What is collected is specific: prompts, code, project files and configuration, the tool calls Bolt makes, edit histories and the fix traces Bolt generates, de-identified and then passed to the company's partner to train open-weight models. Sessions originating in the EEA, the United Kingdom and Switzerland are excluded from training. Access is through Pro, which is $25 a month billed yearly, or through Bolt Lite, a new standalone plan at $9 a month currently behind a waitlist; Forge is not offered on Teams or Enterprise plans. What happens after 14 October is asymmetric and worth knowing before signing up: Lite subscribers keep Forge indefinitely, while Pro subscribers lose it and keep Standard and Max. Extra Forge usage cannot be bought, usage resets on the renewal date with no daily cap, and PDF uploads are not supported yet. On capability, the company publishes its own number rather than a third party's: Forge scores 92.2 on what Bolt calls its Build Index against 101.0 for the best paid model available in the product, a figure measured by Bolt on Bolt's own index.