Privacy Policy

This policy covers the Atlas application, the Atlas marketing site, and the read-only Atlas MCP connector — together “Atlas” — operated by Datacoves, Inc.

Atlas is a separate product with its own documents. This is not the Datacoves platform's privacy policy and does not incorporate it: pointing at a different product's policy would misstate what governs your data here.

Where your organization has signed a data processing agreement with us covering Atlas, that agreement controls for the data it covers, and this policy explains the same processing in plain language.

Atlas is built and operated by Datacoves, Inc., a Delaware corporation with offices at 612 Lynwood St., Thousand Oaks, CA 91360, United States.

Privacy questions and data-subject requests: privacy@datacoves.com. Product and account support: support@datacoves.com.

Atlas holds two kinds of data, and our role differs for each. Getting this distinction right is what tells you who to ask.

• Account and identity data we need to run the service, our billing records, and our own operational logs: Datacoves is the controller.
• Everything your organization puts into Atlas — conversations, agreed definitions and other knowledge base content, tasks, artifacts, connection configuration, uploaded files: your organization is the controller and Datacoves is the processor, acting on its instructions.

In practice: if you are a member of a customer's workspace, a request about workspace content belongs to that organization first. Send it to us anyway if you don't know who to ask — we will route it and help them answer.

• Account and identity: your email address, display name, avatar, your identity provider's subject identifier, your group and role memberships, and sign-in and session records.
• Workspace content: conversation messages, agreed definitions and other knowledge base pages, Canvas artifacts, tasks, comments, uploaded attachments, and the version history and authorship attached to all of it.
• Connection configuration: hostnames, account, database and schema names, and the *names* of the credentials a connection uses. Credential values themselves are stored encrypted and referenced at the point of use — never displayed back, never logged.
• Operational telemetry: request metadata, timings, error traces and background-job records, scrubbed of secrets and content by default.
• Product analytics: metadata-only usage facts — which kind of event happened, when, and in which organization — with no message or document content.
• Audit records: who did what and when, in a tamper-evident append-only log.
• Marketing site and contact form: what you type into the form, plus ordinary server request logs.

• To provide Atlas to your organization and to authenticate you.
• To keep the service secure and available: abuse and fraud investigation, rate limiting, and incident response.
• To support you: diagnosing problems you or your admins report.
• To improve reliability and performance, from aggregated, metadata-only signals.
• To meter and bill under your order form.
• To meet legal obligations and respond to lawful requests.
• To send service notices, which we always send because they are how the service reaches you, and product email, which goes only to people who asked for it and is unsubscribable per category.

We do not use your organization's content to train models, to build advertising profiles, or to enrich any dataset we sell.

For processing where Datacoves is the controller, and to the extent the GDPR or a comparable law applies, we rely on:

• Performance of a contract — providing Atlas to your organization and to you as one of its authorized users.
• Legitimate interests — security, abuse prevention, service reliability, metadata-only product analytics, and establishing or defending legal claims.
• Consent — optional product and marketing email, and anything else we explicitly ask you to opt into. Withdrawable at any time, without affecting your use of Atlas.
• Legal obligation — tax and accounting records, and lawful requests.

For workspace content, your organization is the controller: its own legal basis and its instructions to us govern, and we process only to deliver the service.

Atlas is bring-your-own-key. Your organization supplies its own model-provider credential, so the inference relationship — and the bill — stays between your organization and that provider. Datacoves does not resell model tokens.

Atlas never sends your content to a model provider for training, and adds no training use of its own. Because the credential is your organization's, your agreement with that provider governs what it may do with a request; the major providers' API terms do not train on API content by default.

On a provider whose API retains conversation state by default, Atlas explicitly turns that off. Every request is stateless and the transcript is re-sent from your organization's own database each turn, so prompt and response logging stays inside the Atlas boundary rather than accumulating at the provider.

What leaves on a turn is bounded and visible to you:

• The system prompt, the room transcript, the tool definitions, and the results of tools the agent actually called.
• Every one of those tool calls is recorded in the room's activity trace, so the record of what was sent is yours to read, not only ours.
• What is not sent: credentials, connection secrets, another organization's data, and anything outside the acting person's own permissions.

Open-web research is a per-organization switch, on unless your admins turn it off. While it is on, the search queries the model composes go to the provider's own search tool — capped per turn, and optionally confined to a domain list you control. Those queries are recorded in your audit log.

To make knowledge base search work, page text is sent to Amazon Bedrock to be turned into embedding vectors. Bedrock is a sub-processor, the request is not used for training, and the resulting vectors are stored in your organization's own tenant-scoped rows.

Model output is a suggestion, not a decision. Atlas makes no automated decision about you with legal or similarly significant effect: a definition becomes a record of agreement only when the people who own it sign off.

When you connect an outside AI assistant, that assistant makes read-only JSON-RPC calls to a single Atlas endpoint. Each call carries an OAuth access token bound to one organization, and Atlas re-checks your own permissions on every read.

• What crosses the wire: your search text, and the governed Knowledge Base content you can already read.
• What never crosses it: warehouse data, credentials and connection configuration, content outside your reach, and any write of any kind.
• A page you may not read and a page that does not exist return the identical answer, so the connector cannot be used to discover what exists.

Atlas records that a read happened. It does not store the assistant's side of your conversation, and it has no way to reach into the assistant that called it.

Atlas reads from systems your admins connect, using credentials your organization supplies. Each connection is scoped, and none of them route data to Datacoves for any purpose other than answering in your own workspace.

• Data warehouse: read-only for analysis. The one write Atlas ever performs is into an isolated scratch schema for validating a query, and it is audited and attributed to the person whose turn it was.
• Document sources such as Notion: fetched on demand with your organization's own token, and only for pages the agent has been pointed at.
• Git repositories and metadata platforms: read for context about your models and pipelines.
• Profile pictures: fetched server-side from the URL your identity provider supplies and stored on your organization's own row, so your browser never makes a request to the provider and nothing is shared between organizations.
• External MCP servers: only those your operator has configured. There is no open list, and an unlisted server is refused before any connection is opened.

Every outbound request Atlas makes is checked against a network policy before it is dialled, and redirects are not followed — so a request can only reach the destination it was authorized for.

No third-party analytics or advertising code runs inside the Atlas application. What we measure, we measure ourselves, and we measure metadata.

• Product analytics: an append-only ledger of usage facts, scoped to one organization, recording the shape of an event and never message or document content. It is written out of band, so it can never affect or delay an answer.
• Traces and logs: OpenTelemetry, with deny-by-default scrubbing so secrets and content do not reach a log line.
• Error tracking: scrubbed stack traces and request metadata, so we can fix what broke.
• Cost figures: estimated from token counts at query time for your own visibility. We do not bill per token.

Atlas sets one cookie: the signed session cookie that keeps you logged in. It is strictly necessary, and there are no advertising, cross-site tracking or profiling cookies.

The Atlas marketing site currently runs no third-party analytics or advertising tags and sets no cookies of its own.

We do not sell personal information and do not share it for cross-context behavioural advertising, so there is no ad-tracking opt-out to offer. We honour Global Privacy Control and Do Not Track signals as a matter of course, because we do not track you across sites in the first place.

If that ever changes — for example if we start paid advertising — we will publish a cookie notice and, in the EEA, UK and Switzerland, ask for consent before setting anything non-essential.

For the Datacoves-hosted Atlas service, these providers process data for us:

• Fly.io — application and database hosting, United States.
• Amazon Web Services — object storage for the files you upload and the avatars and exports Atlas generates (Amazon S3), and embedding generation for knowledge base search (Amazon Bedrock). United States.
• SendGrid — transactional email: sign-in codes, invitations and notifications.
• Sentry — scrubbed error traces and request metadata, for diagnosing failures.
• Your organization's chosen model provider — currently Anthropic or OpenAI — reached under your organization's own credential.

A self-hosted Atlas deployment has whatever list its operator configures, and Datacoves is not in that path at all. Where your organization has a data processing agreement with us, notice of sub-processor changes is governed by it; the current list is always available from privacy@datacoves.com.

The Datacoves-hosted Atlas service runs in the United States, as do the sub-processors listed above.

If you are in the EEA, the UK or Switzerland, that is an international transfer. We rely on the European Commission's standard contractual clauses, with the UK international data transfer addendum where it applies, together with supplementary technical measures — encryption in transit, encrypted credentials, and database-enforced tenant isolation.

Where data genuinely cannot leave your own boundary, Atlas can be self-hosted: your infrastructure, your model endpoint, no customer content at Datacoves. Ask us about it.

• Tenant isolation is enforced in the database, not only in application code. Every organization-owned table carries row-level security with a fail-closed policy, and the application connects as a role that cannot bypass it — a query with no organization established raises rather than quietly returning rows.
• Credentials are write-only: encrypted at rest with AES-256-GCM, referenced by name at the point of use, and never rendered back to a user or written to a log. A reference that no longer resolves fails closed instead of falling back.
• All traffic is TLS-only, with HTTP Strict Transport Security.
• The audit log is a per-organization SHA-256 hash chain, so an edit is detectable even by the database owner.
• Outbound network access is deny-by-default: the agent's tools run behind an egress gate that records every call it allowed and every call it refused.
• Access to production is least-privilege and logged.

No system is perfectly secure, and we do not claim otherwise. If you find a vulnerability, tell us at support@datacoves.com and we will work with you on it.

• Workspace content is retained for the life of the account, plus whatever your order form or data processing agreement specifies.
• Deleting a conversation moves it to a 30-day Trash so it can be recovered; a daily sweep then hard-deletes it. That sweep is the deletion — nothing is kept behind it.
• Knowledge base version history is immutable by design, because it is what makes an agreed definition auditable. Retiring a page hides it everywhere and keeps its history until the organization itself is erased.
• Audit records are append-only for the life of the organization: they are the record of who agreed to what.
• Telemetry and error traces are kept for short operational windows and then expire.

Erasing an organization hard-deletes its data, propagated by database-level cascades, audit records included. Atlas hard-deletes by default; the conversation Trash is the one deliberate delay, and it is bounded.

Backups roll off on their own schedule, so a deletion reaches backups when the backup containing it expires.

Depending on where you live, you may have the right to access your personal data, to have it corrected or deleted, to receive a portable copy, to object to or restrict certain processing, to withdraw consent, and to complain to your supervisory authority.

Send requests to privacy@datacoves.com. For workspace content we are the processor, so we will route your request to the organization that controls it and help them respond. We may need to verify your identity before acting, and we will not use what you send us for verification for anything else.

We respond within 45 days, extendable once by a further 45 days where a request is complex, with notice of the reason. There is no charge unless a request is excessive, repetitive or manifestly unfounded — in which case we will tell you why and what it would cost before doing any of it.

Portability is built into the product, not only into this policy: your knowledge base exports as a Markdown archive, and artifacts export as CSV or DOCX. If you leave, the record leaves with you.

We do not sell personal information, and we do not share it for cross-context behavioural advertising. We have not done either in the preceding twelve months.

Categories of personal information we collect:

• Identifiers — name, email address, identity provider subject, IP address.
• Commercial information — your order form, billing and usage records.
• Internet or network activity — request metadata, timings, and in-app usage events.
• Professional information — job title or team, where you or your organization supply it.
• The content you and your organization create in Atlas.

California residents have the right to know, access, delete and correct their personal information, to receive it in a portable form, to limit the use of sensitive personal information, and not to be discriminated against for exercising any of these. The process and timelines are the ones described above; an authorized agent may act for you with proof of authority.

Under California's “Shine the Light” law (Civil Code § 1798.83), we confirm that we do not disclose personal information to third parties for their own direct marketing.

Atlas is a workplace product. It is not directed to anyone under 18, and we do not knowingly collect personal information from them. If you believe a child has provided us data, tell us at privacy@datacoves.com and we will delete it.

When this policy changes we update the date at the top of the page. For a material change we give notice in the app or by email to workspace admins before it takes effect, and continuing to use Atlas after that is acceptance of the updated policy.

Privacy questions and requests: privacy@datacoves.com. Product and account support: support@datacoves.com, or use the contact form.

By post: Datacoves, Inc., 612 Lynwood St., Thousand Oaks, CA 91360, United States.

If you are in the EEA, the UK or Switzerland you may also lodge a complaint with your local supervisory authority, though we would rather you gave us the chance to put it right first.