No index of your company.
It searches your tools when you ask and reads only what the answer needs. Text it reads is cached, encrypted, and purged after 30 days unused. Disconnecting a tool deletes its tokens and its cache.
Chats are different: they stay until you delete them. A deleted chat disappears at once and is purged 30 days later. The one exception is a bot's conversation, deleted with its chat after 12 months without a new message. Chats and the files in them are encrypted at rest. A chat you open stays readable offline on that device, encrypted there with a key that never leaves it, until you sign out or lose access. Code works on unencrypted copies on our job server: a run's files for the run, and a repository a chat copied in until an hour after the chat last used it, a day at most.
No one can quietly read your work.
Admin screens never show chat or document text, and the database stores chats as ciphertext. Deeper production access needs a reason, a recorded session, and a review.
It asks before it changes anything.
Every tool starts at Read. Switch one to Read and change, and each change still waits for you, shown exactly as it will run. You can allow a kind of change for the rest of a chat, or let a chat make edits without asking; deleting, sharing, and access changes still ask, unless your admins allow "Allow everything" and you turn it on in that chat. Every decision is logged, without its contents. Admins choose how far anyone may go, or turn changes off for everyone.
Your computers work the same way. A folder you share starts read-only, you allow changes in it on the computer itself, and each change waits for you here. What a change replaces or removes goes to that computer's trash, never away for good. Commands run only in folders you allow them in on the computer, once your admins turn them on, and each waits for you here too, unless you turned on "Allow everything" in that chat. In your browser it uses only tabs in its own tab group, on the sites you allow on the computer. Of a link it sees only the path on those sites, never its query or a token in it, and nothing of a link to anywhere else. It never reads password, card, or one-time code fields.
Google data, under Google's rules.
At Read, Google Drive gets only Google's read-only scopes. At Read and change it also asks for Google's full Drive scope, so it can create Docs and Sheets, add to or replace text in a Doc, rename, move, and share files, and move them to the trash, each after you approve it. Trashed files stay in Drive for 30 days. Gmail saves drafts and sends the email you approved, word for word, and never labels, archives, or deletes mail. Google Calendar creates, changes, and cancels events and answers invitations, after you approve each one. Each Google product you connect gets its own token, holding only its own scopes.
What we receive from Google APIs is used only to answer you and make the changes you approve. We don't sell it, use it for ads, or train models on it, and our use and transfer of it adheres to the Google API Services User Data Policy, including the Limited Use requirements.
Sharing shares answers, never access.
Each person connects their own accounts. In a project chat, teammates read the answers and see which tool was searched, but only the person whose account returned a result sees it in full. Asking the next question runs with the asker's own tools. Guests read only the chats shown to them, without the links.
In Slack, Microsoft Teams, Discord, or Campfire, a bot answers: each
place answers as the one its owners and admins chose, and in Discord
the one each person picks for their /ask. It answers with
the integrations its team gave it, and writing to it never lends it
yours. Its changes wait for its team in Chat with Work unless they let
it run. Only the answer's text
goes back, and Slack's /cww answers in your own DM with
the app. We check that every request comes from the platform before
reading it, except Campfire's, which Basecamp doesn't sign: a secret
in its address is the only check, so the bot treats what it hears
there as anyone's words. What's posted there is kept by that
platform: Slack, Discord, and Basecamp mostly outside the EU, Teams
where your organization's Microsoft 365 keeps its data.
No one trains models on your data.
The model gets your question, the chat, and what your tools returned. Nothing else. Our model providers' business terms say they don't train on it, and neither do we. The app runs in Germany. Gemini runs in Google Cloud's EU region (a step Google is too busy for goes to Claude on AWS Bedrock in the EU, and the chat says so), GPT in Azure's EU Data Zone, and Claude on AWS Bedrock's EU regions. The same models write chat titles and read scanned PDFs.
Web search stays off until you turn it on.
The assistant can search the web, but only once an admin allows it for your organization and you turn it on in a chat. The searches then go to Google for Gemini or Bing for GPT, outside the EU, so it stays off by default. The assistant is told to keep private details out of what it searches for.
MCP servers you add are yours to trust.
Add any public HTTPS MCP server to your account; its secrets are stored encrypted. It starts at Read, offering only its read-only tools, and every other call waits for approval. What the server does with requests is up to whoever runs it.
Uploads are scanned first.
ClamAV scans every upload before the assistant reads it. If the scan fails, the file is deleted and the answer doesn't run. SVG and Flash files are blocked.
Every release is checked.
Every release passes tests, dependency audits, a Rails security scan, and secret scanning. CASA Tier 2 and Google's verification of our Google scopes are renewed every year.
GDPR, in writing.
Every business customer gets our Data Processing Agreement under Art. 28 GDPR, as part of the Terms of Service: you're the controller, we process for you, and a breach reaches you within 48 hours. The companies that process your data for us are listed with what each receives and where, and you hear 30 days before a new one starts. The services you connect aren't among them: they work under your own agreements with their vendors. Every part, and where it runs, is on one page: GDPR-compliant AI, in writing.
Or run it yourself.
Need it inside your own network? Read the self-hosting details or email hello@chatwithwork.com.
Questions?
For security questions, vendor reviews, CASA Tier 2 evidence, or SOC 2 status, email hello@chatwithwork.com.
Found a vulnerability? Email security@chatwithwork.com, and read how we handle reports: what's in scope, an answer within 3 business days, and safe harbor for good-faith research.
For where your data runs, what's encrypted, and how long it's kept, step by step, read Privacy and EU hosting in the help center.
