Agnoclast — legal

Security

How your data is protected, written for people who read architecture docs. Every claim here is verifiable — several of them you can check yourself from inside the product. Last updated 2026-09-20.

Tenant isolation is enforced in the database, not the app

Every organization is a separate tenant. Isolation is enforced by Postgres Row-Level Security: every read runs as the requesting user, and the database — not application code — decides which rows exist for them. There is no in-product code path where one tenant can query another's data, because the query layer itself is scoped.

One deliberate exception, in the stricter direction: activity from a newly connected source that has no routing rule yet is held outside every tenant rather than filed into one by guesswork. Held records are readable only by the individual whose connected account produced them — narrower than tenant scope, not wider — and become subject to normal tenant and tier rules only once the source is routed to a workspace. This is a separate table rather than an unassigned row inside the tenant tables, so no tenant-scoped policy has to special-case it: a policy cannot return rows from a table it does not reference.

Access inside your org follows your rules

Every page has an owner, and the owner chooses its audience — accessible (org-wide), a part of the org chart relative to the owner (everyone below them, their direct reports, or their peers, each with the owner's managers), scoped (owner + management chain), or confidential (owner + explicit grants only; the management chain is blocked). Captured activity is filed under pages, and a record is visible to whoever can read the page it is filed under (in organizations created before 2026-10-02, from 2026-10-16 unless their owner agreed sooner). Three things stay narrower than their page: a confidential record (owner + explicit grants only), a record its owner has locked (owner + management chain + teammates explicitly added to the same project), and AI-session summaries captured before 2026-10-01, which are treated as locked until their owner releases them. Email and iMessage records are visible only to the people on the message. Confidential never widens. Tiers are enforced by the same database-level policies as tenant isolation, including in search: a keyword that appears only in a confidential page returns nothing for an uncleared viewer — not even a hit count. Counts never widen visibility.

Isolation is re-verified on every code change

A security test suite runs against the production database in CI on every merge to main: cross-tenant leak tests, role-based access tests, tier-visibility tests (including the search-inference case), timeline and page-privacy tests, and operator-access-log tests. A violation fails the build. This is the guard against the failure mode where an app is secure on launch day and drifts afterward.

Raw work content normally stays on your machines

AI-session summarization and page authoring run locally on each member's own machine, under the member's own AI subscription. What reaches our servers is normally the derived summary and the authored page. When local summarization is unavailable, the most recent part of the session — with detected secrets redacted — is sent to our servers so that our AI inference subprocessor can produce the summary. We do not record screens, keystrokes, or browsing, and file capture stores summaries and fingerprints, not file contents.

Operator access: purpose-limited, and only partly recorded

The people who operate Agnoclast hold the infrastructure credentials and can reach the production database directly, with owner rights that bypass the row-level security described above. We use that access to apply schema changes, investigate and fix problems, and handle support requests. Today that includes everyday developer tooling, among it AI coding assistants working on an operator's instructions, not only a separate emergency procedure.

Closing that last gap, so that every privileged session is recorded automatically and shown to you, is the next step for this part of the system. Full detail, including the honest limits of this model: Data Access & Tenant Isolation.

Authentication and encryption

AI processing

Server-side AI processing (connector summarization) goes to Anthropic under API terms that prohibit training on your content. No customer content is used to train models, by us or by our subprocessors. Current subprocessor list: /legal/subprocessors.

Roadmap: operator-blind architecture

For organizations that need "the operator cannot read it even in principle," an operator-blind architecture (content the infrastructure owner cannot decrypt) is an active design track. We publish what's true today rather than claiming it early — see the roadmap note in the data-access policy.

Responsible disclosure

Found a vulnerability? Tell us: theronpeterson654@gmail.com (reference: `/.well-known/security.txt`).

Questions

Security review requests, questionnaires, or anything this page doesn't answer: theronpeterson654@gmail.com.