How it worksSecurity and approval
Security and approval
The rules Mindset enforces for every agent with no setup, how write operations are approved, how credentials are kept and where your data goes.
Skills for your coding agent
Every agent runs under these rules, with nobody setting them up. With the settings your admins choose, they make up your org's compliance posture. If you write your own agent policy, list them as controls the platform enforces.
Why it matters#
No agent ships without these controls, and no agent can grant itself a write. A person, or a setting only an admin controls, decides which writes exist.
What Mindset enforces#
| Rule | What holds |
|---|---|
| Operations are the only way out | An agent, script or function reaches an outside system only through an operation on a connection. There's no general web or HTTP tool, so even a hijacked agent can reach only its operations |
| Writes need consent | Once per operation, from a person or the org setting. No agent can give it |
| Untested writes don't go live | Behavior tests refuse every write, and test runs write only what an admin agreed to. A script or function that calls an operation with no successful call since its last lapse can't go live unless an admin overrides it on the record. See the go-live gate |
| Every action is recorded | Mindset records each model call and operation call in the run's record, and each change in the audit record, so an agent can't skip either. The database refuses deletes from the audit record. See Runs and Audit and personal data |
| Environments can't see each other | An agent in Sandbox can't reach a connection in Production, and an API key works only in the environment it was issued into. See Environments |
Approving writes#
A write operation needs consent before anything can call it. It's given once per operation, never per call, and after that the write runs whenever an agent calls it. A read never needs consent.
| Consent | Meaning | Can an agent call it? |
|---|---|---|
| Not required | A read | Yes, once it's enabled, available and assigned |
| Pending | A write waiting for a person to enable it. The Studio shows Enable it | No. A call is refused, not queued |
| Granted | A person, or the org setting, agreed to it | Yes |
A person's consent is recorded with their name and the time. A write whose definition changes needs consent again.
Your org's default: On (auto-enable)
One org setting decides what happens to a new write: Enable actions that change other systems automatically, under Settings → Systems. Only an org admin can change it.
| Setting | A new write operation |
|---|---|
| On (auto-enable), the default | Is granted as soon as it's registered, so your AI client can enable it with no person involved |
| Off | Waits as Pending until a person enables it on the connection |
Locking and disabling
| Button | On | What it does |
|---|---|---|
| Lock operation | Any operation | Stops every caller. Only a person can enable it again |
| Enabled switch | A whole connection | Turned off, refuses every call to any of its operations before anything leaves |
Each takes effect on the next call, with no new agent version.
An agent can never give consent
No agent can enable a write that's waiting for a person, or enable again an operation a person locked, and the server refuses both from anything but a signed-in person, including an API key. An agent that registers such a write gets a link to the connection's page to pass to an admin.
Consent can't say yes to one supplier query and no to another. For that, a script asks a person before the action and parks the run until they answer.
Live demoHelp widget · LedgerlyA billing help agent reads charges over MCP and stops at a refund nobody has approved.Credentials never reach the agent or the model#
A connection's credential is kept in a secret store and never put into an agent's context or sent to a model. A person enters it on a separate Enter a secret value form, and the model gets back only the field's name. A call aimed at a different host than the connection's is refused before the credential is looked up.
The outside system sees the connection's account, shared or the person's own, so it can't tell which agent called. Mindset's audit names the agent. Scope each credential to what its operations need.
Where your data goes#
| Data | Where it goes |
|---|---|
| Your org's agents, conversations, runs and audit | The Mindset deployment your org lives on. Its region is fixed when the org is created |
| Model calls by your agents | The provider whose model key your org set: Anthropic, OpenAI or AWS Bedrock |
| Model calls by Orca and the builders | Platform-provided keys, unless an admin turns on Run the built-in agents on our own key |
| Model calls with AWS Bedrock | Your AWS account, in the region you name. Some models route across that region's geography, which can include other countries |
| Knowledge search embeddings | OpenAI text-embedding-3-large, on your own embedding key and endpoint if you set them, otherwise the deployment's |
| Outside systems | Only what an operation sends, to the connection's own host |
Personal data can be replaced with placeholders before it reaches the model provider (off by default). See Audit and personal data.
What Mindset doesn't enforce#
- Behavior tests before go-live. Saving and activating don't run them. See Test behavior.
- Cost caps or data residency on an AI model connection. Its limits are the size cap and personal data protection.
- Ownership and housekeeping. Owners, backups and archiving unused agents are your team's conventions. See Roll out Mindset.
Every org setting, with its default, is on Org settings. To approve or withdraw a write, see Approve what agents can do.