Security & your data

Plain answers to the questions teams ask before they trust a tool with their work. Where something isn't in place yet, we say so. Questions we haven't covered? Use our contact form.

Does Teamski train AI on my data?

No. We don't train or fine-tune any model on your messages, files or your agents' memory.

To answer, the relevant part of the conversation is sent to the model that answers it - a built-in model, or a provider whose key you or your team added - and that provider handles it under its own terms. OpenAI, Anthropic and Google's paid API say they don't train on API requests by default. Google's free Gemini API tier may use requests to improve Google's products, so use a paid key for sensitive work. Check each provider's policy for the details.

How long do you keep it?

• Messages, files and agent memory stay until you delete them, or until the project or your account is deleted.
• You can delete messages, connected accounts and keys at any time.
• Deleting your account (Settings → Account) removes your profile, sign-in, keys, connected accounts, the messages you wrote, and any project only you were in. Projects shared with others are handed to another member, so the team keeps its work.
• Copies can remain in backups and logs for a limited time. See the Privacy Policy.

How is it protected?

• In transit: HTTPS everywhere, with browsers told never to use an unencrypted connection (HSTS).
• At rest: the database and uploaded attachments are hosted on Supabase, which encrypts stored data (AES-256).
• A second layer for the sensitive parts: API keys, tokens for connected accounts (Google, GitHub and apps) and your agents' memory are encrypted again with AES-256-GCM, using a key only our server holds.
• Who can see what: every table is protected by row-level security, so people only ever see projects they belong to.
• Files that agents create in a project are stored on our application server.
• Messages and files are not end-to-end encrypted - the AI has to read them to help you.
• Certifications: we don't have SOC 2 or ISO 27001 yet. If your company needs a security review, use Contact sales on the Enterprise plan and we'll answer your questionnaire honestly.

Permissions and oversight

• Three roles per project. Owner: everything, including billing, deleting the project and handing it over. Admin: members, settings, connections and schedules. Member: uses the project.
• Agents ask first. Deleting files, writing to Google Sheets, opening GitHub issues, and any connected-app action that can change something wait for a person to approve them. Background tasks can't do those at all.
• An activity trail for every agent task - what it did, step by step, visible in the project.
• A spend report on the Team plan: who used the project's shared key, and roughly what it cost.
• Not yet: per-channel permissions (everyone in a project can see all of its channels) and a full audit log of admin actions. If you need these, talk to us about Enterprise.

What it connects to

• Built in, on every plan: web search and reading web pages, the project's own files, Google Sheets (read, and add rows), GitHub (read files, search code, list and open issues), and image generation.
• Connected apps, on Team: Notion, Linear, Jira & Confluence, Asana, Sentry, Stripe, Canva, Hugging Face and DeepWiki, through each app's official MCP server - plus any other MCP server by its address. Agents can use whatever the app's server offers; anything that changes data waits for approval.
• AI models: the built-in models on every plan, a free Google Gemini key on any plan, and on Team your own or a shared key for Anthropic (Claude), OpenAI, xAI (Grok), Groq, Cerebras, NVIDIA, DeepSeek, Mistral, OpenRouter, Together AI, or a model server you run yourself.

How reliable are long-running agent tasks?

• Background tasks run on a separate worker, not in your browser - close the tab and they keep going. You're notified when one finishes or fails.
• If our server restarts mid-task, the task is picked back up from the step it had reached, not started over.
• You can pause or stop a task at any time, even mid-step.
• If a connected tool fails, the step is retried without it rather than failing the whole task.
• The limit to know: a single task works in up to 20 steps, and the last step always writes up a result. For very big jobs, split the work into several tasks, or put it on a schedule.
• A scheduled agent that is still busy with its previous task waits up to two hours, then skips that run instead of piling up.

Can I take my data with me?

Yes. A project's owner or admins can export the whole project from Settings → People → Export: one file with its members, channels, every message, what the agents remember, schedules, and the files the agents wrote. Attachments are listed in the export and can be downloaded from the chat. You can delete your account at any time from Settings → Account.

How good is the AI on the free plan?

• Free: GPT-OSS 120B - OpenAI's open-weight model - served by Groq, with 20 built-in AI messages per person per day.
• When the shared free capacity is busy, it steps down to GPT-OSS 20B, then Qwen3 27B, and the reply says when a different model answered. If all of those are busy, a small model on our own server answers.
• Team: 100 built-in messages per person per day, plus your own or a shared key for Claude, GPT, Gemini and others. Messages on a key are billed by that provider and don't count toward the daily allowance.
• For heavy or sensitive work, adding a paid key of your own is the surest way to get consistent quality.