Bolt Forge Privacy: What Data Is Shared, What Gets Trained and How Opt-In Works

01

Overview

Bolt Forge has a different privacy model from Bolt.new’s Standard and Max agents, and that difference is central to how the research preview works.

Forge gives eligible users a much larger experimental usage allocation, but users who enter Forge explicitly opt in to let anonymized build-session data contribute to training future open-weight models with Arcee AI.

The key point is that this is not presented by Bolt as automatic training across every Bolt project. The data-sharing arrangement is tied specifically to Forge.

02

Does Bolt Forge train on your code?

Bolt says Forge sessions can be used to help train future open-weight models after the user opts in.

The shared session data can include prompts, code and fix traces created during the building process.

That does not mean every piece of code in a Bolt account is automatically used. Bolt states that Standard and Max are outside the Forge training workflow and that users enter Forge through a separate consent step.

03

Is the training opt-in or automatic?

Bolt describes Forge training as opt-in.

A consent screen appears when a user switches into Forge. Bolt says this happens each time rather than relying on a one-time consent buried in account terms.

If the user declines, they remain in Standard or Max and nothing from that switch is shared through Forge.

If the user accepts and builds in Forge, the session can enter the Forge data pipeline.

04

What data can be included?

According to Bolt’s launch documentation, Forge data can include:

• prompts;
• code;
• fix traces generated while resolving problems during a build.

Those traces are potentially valuable because they show not only the final code but how a user and agent move from an error or incomplete implementation toward a working result.

Bolt’s rationale is that real build trajectories can teach future coding models patterns that static public repositories do not capture.

05

What does Bolt say it removes?

Bolt says secrets and personal information are stripped before Forge data leaves its infrastructure and that sessions are anonymized.

The company also says it validates the process against seeded test data.

Those are vendor statements about Bolt’s own pipeline. Public launch material does not provide an independent audit demonstrating the effectiveness of every anonymization step, so users with strict confidentiality requirements should evaluate the arrangement according to their own risk and compliance needs.

06

What role does Arcee AI play?

Arcee AI is the open-model lab working with Bolt on the training side of Forge.

Bolt says anonymized Forge data is transferred to Arcee under a signed data-processing agreement. The goal is to use real software-building sessions to help train open-weight models.

Bolt specifically says the partnership is intended to contribute to a trillion-parameter-class open-weight model and that the resulting weights will be published.

Those future model details should be treated as a stated roadmap until the training run is complete and the weights are actually released.

07

Are Standard and Max sessions used for Forge training?

Bolt says no.

Standard and Max remain separate agents with their existing privacy behavior. Simply being a Bolt customer or switching between Standard, Max and Forge does not mean all work becomes Forge training data.

The training arrangement begins when the user explicitly enters Forge and accepts its consent flow.

That separation is one of the most important facts for anyone deciding whether to experiment with Forge.

08

Are Teams and Enterprise included?

No at launch.

Bolt says Teams and Enterprise workspaces are excluded from Forge and from Forge data collection during the September 2026 research preview.

This also means Forge is not currently an enterprise privacy option simply because Bolt has enterprise products elsewhere in its platform.

09

Can you stop sharing later?

Yes for new sessions: switch back to Standard or Max and you stop creating new Forge training sessions.

But Bolt makes an important caveat. Data that has already been used to train a model cannot be removed from that trained model after the fact.

That makes the decision to enter Forge different from a reversible storage preference. A user can stop future sharing, but cannot assume previously trained-on data can later be extracted from model weights.

10

Does the 50X usage depend on data sharing?

Forge’s launch economics are built around the opt-in research model.

Bolt gives individual Pro users up to 50X more Forge usage at no extra cost through October 14, 2026 and explicitly frames the shared training data as part of the exchange that makes the research preview possible.

Users who do not want to participate can stay on Standard or Max rather than enter Forge.

The exact commercial structure after October 14 is not yet defined.

11

Is Bolt Forge safe for sensitive proprietary code?

There is no universal answer.

Bolt says it strips secrets and personal information, anonymizes sessions and uses a signed data-processing agreement with Arcee AI. Those controls reduce risk, but they do not automatically satisfy every company’s confidentiality, contractual, regulatory or security requirements.

For highly sensitive proprietary code, regulated data or client-confidential projects, users should not rely on a marketing summary alone. Review Bolt’s current legal and privacy documentation and your own obligations before opting in.

The simplest operational rule is: if a project should not participate in an experimental training-data program, keep that work in Standard or Max rather than Forge.

12

Does anonymized mean fully risk-free?

No privacy system should be treated as risk-free merely because data is described as anonymized.

Bolt’s launch page says secrets and personal information are stripped and sessions are anonymized, but it does not publicly provide enough detail to independently measure residual re-identification or proprietary-code risk.

That does not mean the controls are ineffective. It means the strongest defensible statement is that Bolt says it applies these protections; independent verification is limited in the public launch material.

13

What happens to the resulting models?

Bolt says the data will help train open-weight models and that the resulting model weights will be published.

That is an important distinction from a closed internal model. Users who contribute through Forge are helping create models that Bolt says will eventually be downloadable and inspectable by the wider ecosystem.

The first training run is expected to begin in October 2026. Until those weights are actually released, the publication promise remains forward-looking.

14

Bottom line

Bolt Forge’s privacy model is deliberately explicit: enter Forge, see the consent screen, opt in, and let anonymized build sessions contribute to open-model training in exchange for much larger experimental usage.

Standard and Max remain outside that Forge training arrangement, and switching back stops new Forge sharing.

For casual prototypes and non-sensitive experiments, that trade may be attractive. For proprietary or regulated work, the decision deserves a stricter review because data already incorporated into model training cannot later be pulled back out of the resulting model.