Agnoclast

Data Access & Tenant Isolation

Plain-language summary of how your data is separated and who can access it. Last updated 2026-10-02.

Your data is sealed to your organization

Every organization on Agnoclast is a separate tenant. Your records, pages, projects, and the intelligence built from them are isolated at the database layer — not just in the UI. Access is enforced by Postgres Row-Level Security (RLS): every read is scoped to the viewer's organization and their permitted visibility within it. There is no in-product path for one customer to see another customer's data, in either direction. This isolation is verified by an automated security test suite that runs on every code change.

Inside your organization

Visibility is decided by pages. Every page has an owner, and the owner chooses its audience:

You set the audience of pages you own. Owners and managers grant within what they can already see; grants are logged and revocable.

Who can see captured activity

Captured activity — commits, meetings, AI-session summaries and other records — is filed under pages, and a record is visible to whoever can read the page it is filed under. This is deliberate: the person responsible for a page should be able to read the work filed on it, including work done by people they do not manage.

Changed 2026-10-02. Before this change a scoped record stayed with its owner's management chain and teammates explicitly added to the same project, even when it was filed under a page the whole team could read. The change widens what teammates can see, so it is a material change: it applies from the outset to organizations created on or after 2026-10-02, and from 2026-10-16 to organizations created earlier — 14 days after their owner was notified — unless the owner agreed to an earlier date. Until then the earlier rule applies in those organizations.

Three things stay narrower than their page:

Separately, and as before, a scoped record that belongs to a project is visible to teammates who have been explicitly added to that project, whatever page it is filed under. Project membership is a deliberate act: it is never inferred from sharing a repository, channel or folder, and no one is added to a project automatically.

Email and iMessage records are visible only to the people on the message, and calendar events also to the people invited. Activity not yet filed under any page is visible only to the person it belongs to, or to the organization's owner and admins when it belongs to no one. You can lock or release any record of yours at any time.

Connected-source activity waiting to be assigned

Records from a connected source (email, calendar, chat, files) are filed into a specific workspace based on an explicit rule for that source — which account and which source goes to which workspace. That rule is deterministic and set deliberately, so a given source always lands in the same place.

When a source has no rule yet — typically a newly connected one — its records are held separately instead of being filed into a workspace by guesswork. While held:

We do this because filing activity into the wrong workspace is effectively permanent — records de-duplicate per workspace, so a source split across two workspaces cannot be cleanly reassembled. Holding an unassigned record is recoverable; misfiling it is not. The individual can see what is waiting, and the operator-access controls below apply to held records exactly as they do to filed ones.

What the operator can access — and the controls

Agnoclast is operated software, so — like every hosted product — the people who run it hold the infrastructure keys and can reach the database directly, with owner rights that bypass row-level security. We don't hide that. Here is how it works today:

We deliberately do not claim “the operator is cryptographically unable to read your data.” For a product whose value is reading your work to summarize and connect it, that claim would be untrue today. We claim what's real: access is purpose-limited, contractually bound, and partly recorded and visible to you, and we say plainly where it isn't recorded yet. Recording every privileged session automatically, and showing it to you, is the next step.

Where your content is processed

To build summaries, search, and shared context, content is processed by our infrastructure and inference subprocessors (e.g. the LLM provider for summarization, the database/host for storage). These are operational processors under contract; content is not used to train third-party models. Where possible, AI processing runs on your own machine under your own AI subscription — session summaries and page authoring never send raw transcripts to our servers. A current subprocessor list is available on request.

Roadmap — operator-blind architecture

For customers with regulated or contractual zero-access requirements, an operator-blind model (content the operator cannot read even with infrastructure access) is on our roadmap as an enterprise milestone. It changes the architecture deliberately; we'll scope it with you if you need it.

Questions or a security review request: contact the Agnoclast team.