Your work stays on your {{device}}.

Your prompts travel. The whole map on one page: what stays, what leaves, and what we keep.

What we don't do.

  • We don't read your conversations
  • We don't train models on your data
  • We don't sell or share your information with third parties
  • We don't store conversation content on our servers
  • Your agent reaches local files with no cloud connection
  • Your agent's model-catalog refresh and its startup update check are both switched off
  • Your sandbox blocks the plugin-catalog request by default
  • No favicon fetching, no location requests, no triggers running until you set them up

What one request does.

Your team works on your machine. When it needs to think, one request goes out and the answer comes back.

On your {{device}}

  • Conversations
  • Files
  • Team memory
  • Agent configuration
  • Workspace data

We can't see, access, or retrieve any of it

Maxiful

Model proxy: authentication and usage management. We don't store the content of the call.

What we keep: your account, billing, and usage metadata.

Model provider

No training on customer data. Kept only briefly, for operations and abuse monitoring.

Every item, in two columns.

Stays on your {{device}}

  • Conversations: everything you say to your team and everything it says back
  • Files your agents read, write, or process
  • Team memory: what they have learned about your preferences, patterns, and priorities
  • Agent configuration, skills, and personalisation
  • Workspace data: emails fetched, calendar entries cached, research compiled

Unless you grant a time-limited support session, we can't see, access, or retrieve any of it.

Leaves your {{device}}

  • Prompts, routed through our proxy to the model provider: your message, the relevant conversation history, the task instructions, and excerpts of your local data when the task needs them
  • Account and billing: name, email address, plan tier. Stripe holds the card, and we don't see the full number
  • Usage metadata: model name and token counts, buffered up to 7 days. Not content
  • Service metadata: deployment status, agent health checks, error logs. No conversation content
  • Diagnostics: whether setup finished, which model answered, which screen you were on, which marketplace item you installed. Never message, prompt, or file content, and you can turn them off

The minimum needed to run the service.

The thinking happens in the cloud.

An AI team has to reason, and that runs on models far too large for a laptop. So the prompt leaves. This is how every cloud AI works. The difference is that we draw you the map.

  • Routed, not stored. Calls pass through our infrastructure for authentication and usage management. We don't store their content.
  • No training, on any route. We use only providers that don't train on customer data. Cheaper routes exist through providers with weaker policies, and we don't take them: when privacy and price pull apart, we pay the difference.
  • Metadata, not content. Model name and token counts, buffered up to 7 days, so your plan budget and your dashboard work.

The rest of it, in short.

  • No standing accessWe can't reach your {{device}}. If support needs to see your agent's configuration or logs, we ask first, and only for a time-limited session.
  • If you leaveYour local data stays on your {{device}}, whether you uninstall or not. We delete your account information within 30 days. Billing records are kept for 6 years, as the law requires.
  • Where it runsOur cloud services and CDN run on Cloudflare's global network, and we operate no data centres of our own. Some providers sit outside the UK and EEA; where personal data is transferred, we rely on Standard Contractual Clauses.
  • On your {{device}}Credentials are written with restricted file permissions, and stored in your operating system's secure credential store where available. Skills run with scoped permissions, and high-impact actions wait for your approval.
  • Welcome emailWhen setup finishes, your email address goes once to our welcome-email service, which sends you a getting-started email.
  • Background work that spends your AI budgetThree features run on their own, without a prompt from you, and each is on by default. Self-learning reviews substantial conversations and may create or refine workspace skills, then reviews your whole skill collection once a week (off switch: skills.workshop.autonomous.mode). A session observer writes a short status note after each session, using the cheapest model in your configuration (off switch: gateway.controlUi.sessionObserver). Memory dreaming is set to run a nightly consolidation pass at 03:00 local time (off switch: plugins.entries.memory-core.config.dreaming.enabled); see Local memory search below for how it runs on your own machine. All three call the same model provider as your normal conversations and are billed the same way. No new destination for your data.
  • Personal recallPersonal recall is on by default on an install used by one person, so your agent can draw on your other private conversations with it when it answers you. See Local memory search below for how that recall runs on your own machine. On an install several people share with the same agent, it's off by default, so one member's conversations don't surface in another's.
  • A few messages still say OpenClawMaxiful runs on OpenClaw, and a handful of failure messages your agent posts into chat, for example about a full disk or a broken credential store, currently name OpenClaw instead of Maxiful. We've accepted that for the beta.
  • Local memory searchYour install downloads a 340 MB local embedding model once, so your agent's memory is searchable on your own machine. On most machines, memory content never leaves it. On machines with less than 8 GB of memory or 2 GB of free disk, memory search runs through the same model provider that already answers your agent's conversations, and the app tells you when that is the case.
  • Web searchYour agent's search queries go to OpenRouter and its search partner and are billed to your existing OpenRouter allowance, a fraction of a cent per search.
  • Inviting a colleagueWhen you invite a colleague by email, their email address is sent to Maxiful so the invitation can be delivered, and their membership is recorded there. No conversation, no file and no other detail about them leaves the machine. A colleague invited through your team's own chat workspace has nothing sent at all.
  • Shared-machine memory indexingIf the built-in memory index cannot run on a machine that several people use, conversations, including your colleagues', are sent for indexing to the same provider that already answers them.
  • An unrecognised deviceA device of a colleague that Maxiful does not recognise waits for the owner's approval. Nothing is sent to Maxiful for that; the approval happens on the owner's machine.
  • Connecting another assistantAn assistant you connect can ask your assistant about anything it knows, on this machine; nothing about it is sent to Maxiful.

The formal version.

Our privacy policy covers the same ground in legal language.

Read the privacy policy See what runs where

Maxiful is a product of PolyTrader Ltd (England & Wales), a data controller under UK and EU GDPR, with a designated EU representative under Article 27.