Is your data used to train models?
On an organization account, never. On a personal account, only if you leave the setting on. What is collected, what is never collected, and what deletion can and cannot undo.
On an organization account, never. On a personal account, only if you leave the setting on. Everything below is the detail behind those two sentences.
Personal accounts
There are two settings, both in Privacy controls, and both are on by default:
- How a session went. Metadata about a task: whether it succeeded, how long it took, what it cost. This table has no column that can hold code, prompts, diffs, file paths, or tool arguments.
- The session content itself. The text of a session, kept as a sample for post-training. For a session run in the CLI this is what puts the work on Dropstone's servers: the task, the change the agent made, and the tool output around it. The CLI's own history stays on your machine either way.
Turn either one off and nothing new is collected from that point on. What was collected before stays until you delete it.
Organization accounts
Excluded entirely. The organization's own telemetry setting is off by default, and it is enforced where data is collected rather than in the client, so a member's personal preference cannot turn it on for an organization.
The organization's Data page states it in full: with the setting off, "Nothing from this organization is collected, and no code or conversation from it is ever used for training."
The conversation archive an organization keeps for review is a separate thing, held under the organization's own policy, and it is not fed to Dropstone. See Read and export the conversation archive.
The other half of that: the archive exports as JSONL, and an organization is free to fine-tune its own models on it. Dropstone may not, and does not. See How Dropstone compounds in an organization.
Incognito chats
Never collected for training, on any account, and never saved to your history. See Incognito chats.
One exception, and it is not about training: in an organization that has content logging set to full, a temporary chat is recorded in the organization's own archive like any other conversation. The temporary chat says so itself: "Not saved to your history, but the organization records conversations on this account."
What is not done
- Nothing is graded or scored. Dropstone does not rate sessions, label them, or build a dataset out of them beyond the samples described above.
- Nothing is kept from a file you attach beyond what the conversation needs. See What happens to the files you upload in a chat.
- Nothing is collected from an organization, at any setting.
What a sample contains
A sample is a session rather than a transcript of everything: the task, the change the agent produced, the tool output around it, and three signals about how it ended. Whether the edit was accepted, whether the tests passed, and whether the task resolved.
Those signals are read from the session's own output, not decided by a separate judge, and they are the reason a session is useful as a sample: without knowing whether the work landed, a session is only a transcript.
Whether the task resolved is also what classifies the sample. A session that landed and a session that failed are different things to learn from, and a failure carries a short label for what kind it was: it did not compile, it failed its tests, it ran out of time, or it ran out of context.
Ratings are kept separately, one for each surface: a rating you leave on a reply in chat, and a rating on a session in the CLI. A rating is something you chose to give rather than something collected quietly, and it carries the reply it is about so that it means something later.
Deleting it
If you never opted in, there is nothing to delete.
If you did, the samples are removed with your account, along with your usage records. What deletion cannot undo is the model: anything already used to improve one has been learned from, and that is not reversible by removing the sample afterwards. Turn the settings off before you share something you would not want used.
Related articles
- Privacy controlsThe privacy page in your Dropstone dashboard explains what happens to your data and gives you the switches that control it, including opting out of having your conversations used to improve Dropstone.
- Export and delete your dataHow to get a copy of everything Dropstone holds for you, and how to remove it. Conversation export, memory, project files, CLI history, and account deletion, in one place.
- How an organization manages its dataThe three settings an owner controls for an organization, what each one stores, who can read it, and the limits of the strongest setting. Content logging, product telemetry, and retention, in one place.
- Incognito chatsAn incognito chat writes nothing to your memory: nothing you say or correct is kept, and no summary is made. What Dropstone already knows about you is still available to it, so the work gets the same assistant without leaving a trace.
- How Dropstone handles your dataWhat Dropstone stores, for how long, who can see it, whether it is used to improve the models, and the commitments that apply to every account. A plain-language summary of Dropstone's data practices.