AI and your data
The assistant and the agents work on your base's records the way your team does, under the same rules. This page says what they may see and do, what waits for a person, and how the AI models are reached.
What the AI may do
- Your permissions, no more.
- The assistant, and an agent you chat with, can do only what you may do: the same records, the same fields, the same actions.
- It listens only where it is present.
- The assistant reads the messages of a chat only where it is present: in each public channel of the base, in a thread where it answered a mention, and in a private channel, a group chat or a direct message where one of its users added it or an agent of the base. There it can answer a message that does not mention it, with the permissions of the person who wrote that message; in a public channel its answer uses only the words of that person. It proposes a change, or a record from what a person wrote, only on a card that this person applies. The chat says who added it, any user of the chat can remove it, who manages a public channel can remove it from that channel, and an admin can turn the listening off for the whole base in Settings, AI settings.
- An agent on its own is a user.
- An agent that works by itself, on an automation or a step of a workflow, is a user of the base with a role of its own: it sees and changes only the records and fields that role allows, as a person would.
What waits for a person
- A card for anything large.
- Deleting, changing many records at once and every change to how the base works come as a card, and nothing happens until a person approves it.
- Nothing large unwatched.
- An agent working on its own, where nobody is watching, cannot do any of the things that need a person's approval.
Your records
- One switch.
- In a base's settings, one switch keeps its records, its chat and its files away from the assistant and from every agent. With the switch on, the assistant can still change how the base works.
- Every change on record.
- Every change the assistant or an agent makes is in the history, marked as theirs, and can be undone.
- You know when it is AI.
- The assistant and each agent are named where they speak, the assistant with an AI mark, and the history marks every change they make, as the EU AI Act (Article 50) asks.
The models
- We choose the model.
- We choose the model that suits each task. You need no AI account of your own. What the assistant and the agents send to the AI goes to OpenRouter, in the USA. We cannot name the model hosts, which are OpenRouter’s subprocessors: OpenRouter picks the host for each request among those that neither train on it nor keep it.
- No training.
- Base Captain trains no model on your data.
- Agents on your own machine.
- An agent on your own computer runs on your own AI subscription, such as Claude Code or Codex: what it reads goes to that provider, under your terms with them. Other agent tools on your computer use the AI providers you set up in them, or ours.
- Free models.
- opencode, an agent tool on your computer, also offers free models, marked "free". The provider of a free model can keep what the agent reads and use it to train models.
OpenRouter and the other companies that handle your data for us are on the subprocessors page.
Cost and what is kept
- The AI is in the plan.
- The assistant, the agents and the workflows of your base use the AI that the plan includes, for the normal use of a base (the terms' fair use). An agent on your own machine runs on your own subscription and uses none of it.
- Conversations are kept.
- Nothing deletes a conversation on its own: the base keeps it. You can delete your conversation with the assistant, one part at a time, and the summary of a part goes with it.
- A fault of Base Captain.
- When the assistant finds that Base Captain itself fails and it cannot fix this in your base, it tells you and reports the fault to us: what fails and what it checked, in its own words from yours. The AI agents that we use to build Base Captain read the report to repair the fault. A base that keeps its data from the assistant and agents sends them no report.
- Report this answer.
- Shares that one conversation with us, or for the assistant that one part with its summary, with your note, only when you choose to.