Skip to content
Cognolo

Security and trust

An assistant that can act needs limits that are built in, not promised. This page says what the design does, and just as plainly what has not been done.

Six rules built into the runtime

  1. 1

    Only merged tools can be called.

    The assistant can do what a developer has reviewed and merged, and nothing else.

  2. 2

    It acts as the signed-in person.

    Each person's assistant gets only the tools their role allows, and every permission check your product already has still applies.

  3. 3

    Changes wait for a person.

    Lookups run at once. Changes become a preview card to approve or decline. Actions that can't be undone are flagged and never run automatically by default.

  4. 4

    Your records are data, not instructions.

    Anything the assistant reads from your records reaches the model labelled as data, with a note never to follow instructions inside it.

  5. 5

    Its limits come signed.

    Your app takes rules and trust levels only from Cognolo's signed bundle, for the right environment and never an older one. If the rules can't be checked, the assistant does less, never more.

  6. 6

    Everything is recorded.

    Each lookup, proposal, approval and decline is logged as the assistant's, separate from what the person did by hand.

Trust is earned one skill at a time.

Each skill has its own setting. Move it and see what the assistant would be allowed to do.

Skill: check out a booking

Write
How much the assistant may do by itself

The assistant shows a preview card. A person approves or declines each change.

Planned. A skill would move up only after a run of approvals with no edits (the working number is 25).

What Cognolo reads

It reads your repository, and writes only if you let it.

The GitHub App only reads, unless the owner lets it open pull requests. Even then it only makes a new branch for its own proposal, and never merges: a developer reviews and merges. It is built and tested against a stand-in for GitHub.

Secrets are skipped, and nothing is run.

The code reader never opens .env files, keys or certificates, and never runs, installs or loads the code it reads. A test fails if either rule is broken.

Nothing is callable until merged.

The assistant can use only the tools a developer reviewed and deployed. A tool that was proposed and never merged does not exist to it.

Where your data goes

Stays in your app
Your credentials and your database access. Tools run inside your app, as the signed-in person, so your own permission checks still apply.
Reaches the model
What the assistant looks up to answer someone is sent to the model as part of that conversation. The model is Claude, with your key or ours.
Your customers’ words
Designed to stay in your app. Cognolo counts what the assistant did, never what people typed to it. A conversation filter changes that only if you switch it on, per environment, and even then only after the assistant has turned down a clearly off-topic request: from then, that person’s messages, with the end of the assistant’s reply before each, are checked by Cognolo and TypeSafe’s model before the assistant sees them. Checking ends early after three messages in a row that are clearly on topic (once per conversation), and never lasts more than 30 minutes after the assistant’s latest decline. Nobody else’s messages are sent to Cognolo or TypeSafe. TypeSafe says it doesn’t use what it checks to train its models, but it hasn’t promised to delete it straight away. With a click, a person can send the business a short request the assistant wrote and they saw first; their own words are not sent, and it is kept until the owner deals with it, at most 30 days. For a message it stops, Cognolo counts which everyday topic it was about (a label such as “weather”, never the words), and the person can click to ask the business to allow that topic, which sends only the label. Keeping stopped messages is a separate choice. They are deleted after 30 days, and a message Cognolo detects contains personal details is never kept. The filter is built and has been tried in one real app, and it will stay off for customers until TypeSafe confirms in writing how long it keeps what it checks.
Planned, not decided
The records Cognolo keeps in its cloud, how long they are kept, and who at Cognolo can see them. These are open questions, and they will be written down here before anyone is asked to depend on them.

What has not been done

A list of what is missing is part of trust, so here it is.

  • No security certification of any kind, and no independent security review or penetration test.
  • Only small tests against the real model: 10 graded requests to the assistant (14 November 2026) and 14 cases of turning answers into rules (19 November 2026). Everything else is covered by automated tests that use a scripted stand-in for the model.
  • No hosted service. The cloud, dashboard and billing do not run as a service, and the code reader and GitHub App have not been used on a customer's code, so nothing here describes a running system.
  • Nothing on this site is a recording of the real assistant. The pictures and the demo are illustrations with invented data.

Have a security question we should answer here? Tell us when you ask for early access.

Request early access