Security best practices

The assistant is a powerful tool. Powerful tools deserve thoughtful use. These are habits for getting work done on Linux without handing over more than you meant to.

Be intentional about what you share

The assistant remembers what you tell it. Facts and preferences can be stored as memories and included in later model calls when they are relevant.

Share freely for things that help: projects, preferences, how you like to work. Think twice before sharing passwords, financial account details, medical records, or legal matters. If it is in the conversation, it may end up in a memory and in a future model call.

• Ask it to forget: “Forget what I told you about [topic].”
• Use a private conversation for sensitive topics so memories stay isolated.
• Edit USER.md to remove anything you do not want persisted.

A secret scanner catches accidentally pasted keys in common formats. Treat it as a safety net, not a strategy.

Review workspace and trust rules

State lives in files you can read. Every few weeks, glance through the hatch workspace (under ~/.local/share/cadence/assistants/ on Linux).

• USER.md: still accurate?
• SOUL.md: still the behavior you want?
• IDENTITY.md: still happy with the name and personality you chose?

Also review Settings → Permissions & Privacy. Trust rules accumulate. You may have approved something broadly that you would rather scope.

Understand what you are approving

1. Read the prompt. Do not click Allow out of habit.
2. Check the scope: one file, or the home directory?
3. Look at the risk level.
4. Does this action make sense for what you just asked?
5. Allow can create a trust rule for similar future actions. Scope it if you want more control.
6. Say no if you are unsure. The assistant will not retry automatically.

It is easy to start clicking Allow reflexively. Each prompt is a new action. Take the half-second.

Be cautious with custom skills

Custom skills can read files, run commands, and make network requests. Treat them like any software you install.

• Ask “Show me what this skill does” before you save it.
• Third-party skill tools are prompted by default. Writing skill source is high-risk.
• If you install from a community catalog such as skills.sh, read SKILL.md first.
• Run a new skill with one-time approvals before creating persistent trust rules.

Credential hygiene

Credentials sit behind the Credential Execution Service (CES), a separate process, so the assistant does not see plaintext keys in the model context. On a local hatch they live in an encrypted store for that instance. There is no Aevum-hosted CES in 0.1.

• Grant the minimum OAuth scopes. Read-only when possible.
• Rotate stored API keys every few months.
• Ask “Show me my credentials” and delete stale ones.
• Do not store master passwords in the vault. Tokens and API keys, not your bank login.

Computer use and host tools

macOS Computer Use (GUI click, type, OCR) is not the Linux 0.1 path. On Linux, host access is host_file_* and host_bash after you grant it, plus the Browser skill for the web. Screenshots and page extracts still go to the model you configured. Do not leave banking or medical screens in view if a screenshot is about to be taken.

Detail: Computer use.

Channel security

Telegram, Slack, phone, and other channels use the same verification idea. When you add a channel, you complete a handshake (a code in the authenticated desktop session, entered on the new channel). Until that handshake, the assistant does not talk on that channel.

If someone else contacts the assistant, you get a chance to add them as a trusted contact. Trusted contacts can chat. Sensitive actions still need your approval. Unverified strangers get a hard-coded refusal, not a model-generated one.

See Channels and Contacts.

Network awareness

Two kinds of outbound HTTPS: model calls (prompts and context to the provider you configured: Hive, xAI, or another OpenAI-compatible endpoint), and service APIs (mail, calendar, search, browser). Cadence does not default to a hosted Anthropic account in 0.1. You bring the keys. Encrypted in transit does not mean the data stays on the laptop.

Starting over

There is no Cadence Cloud reset button. On Linux:

1. Stop the assistant (cadence sleep / quit the app).
2. Remove or archive that hatch under ~/.local/share/cadence/assistants/.
3. Revoke OAuth apps at Google, Slack, GitHub, and any other vendor you connected.

That is irreversible unless you archived first. Related: Permissions, Privacy and data.