> ## Documentation Index
> Fetch the complete documentation index at: https://docs.odella.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Tool and browser boundaries

> Understand enforced tool policies, isolated workspaces, browser limits, and where approvals do—and do not—apply.

A worker's instructions describe how it should work. **Access rules and action policies determine what it is allowed to do.** Use both, but do not treat a written instruction as an enforced security boundary.

## What deterministic governance means

For supported connected tools, Odella uses explicit policy evaluation to allow an action, block it, or require approval. The worker cannot grant itself permission by saying an action is safe.

The decision uses the resolved tool, its classification, configured policy, and saved exceptions. Unknown connected-tool actions ordinarily require approval under the recommended policy. Explicit blocks take priority over allow exceptions; credential and administration actions classified as credential operations are blocked by this policy.

**Deterministic does not mean infallible.** Tool classification is not proof of what a third-party service actually does. Review the exact tool and provider permissions, especially for [custom integrations](https://docs.odella.ai/integrations/custom-integrations.md).

## Separate the control layers

| Control | What it limits | What it does not replace |
| - | - | - |
| Organization capabilities and resource access | Who can use a worker or perform an operation on a resource. | Provider-side permissions or action review. |
| Worker Governance | Manager, primary audience, permitted editors, and conversation sharing. | Inbound-channel rules or connected-tool policy. |
| Worker resource assignments | Which knowledge, workflows, skills, and tools are available for its responsibilities. | Permission to perform every action exposed by a connected service. |
| Connected-tool policy | Whether supported connected-tool operations proceed, are blocked, or need review. | A universal policy for every browser click or shell command. |
| Outbound messaging policy | Whether supported outbound messages to a recipient proceed, are blocked, or need review. | All possible data transfers through every website or integration. |
| Provider permissions | Which records and operations the connected account can access at the provider. | Odella's worker and resource access rules. |
| Execution isolation | Separation of the worker's work environment from the application and other work environments. | Protection against every permitted use of data already available to the worker. |

## Isolated execution and zero-trust principles

Treat the worker's execution environment as a limited participant, not as a trusted administrator. A running AI Employee does not inherit unrestricted access to the organization. Platform requests carry identity and account context, and connected applications and model access use controlled service boundaries.

Odella separates worker execution from the public application. The Kubernetes worker configuration includes non-root execution, a read-only system filesystem, disabled privilege escalation, dropped Linux privileges, and resource limits. Writable workspace and temporary areas remain available for legitimate work.

These are complementary boundaries: a workspace is not the authorization authority, and a worker's request is not itself permission to access a platform resource. Model-provider credentials are handled at a controlled service boundary rather than provisioned as direct provider keys in the worker's workspace.

> [!WARNING]
>
> “Isolated” does not mean “no network access,” and “zero trust” does not mean
> that every environment has identical network controls. Do not assume universal
> zero-egress execution, a configurable website allowlist, or complete data-loss
> prevention. Contact Odella to verify deployment-specific requirements during a
> security review.

## Connected-tool approvals

Open **AI Employee → Settings → Security & Privacy** to choose a tool policy and review saved rules.

* Use **Approval-first** during a sensitive pilot so ordinary connected reads and writes require review.
* Use **Restricted** when ordinary connected writes should be blocked rather than offered for routine approval.
* Review saved **Always allow** exceptions; they can change the behavior you expect from the default policy.
* Use explicit **Always block** rules for tools the worker must not use.
* Prefer provider accounts with narrow permissions. An approved tool can still have broad access at the provider.

Draft creation and some low-risk organization actions are classified separately from ordinary writes under the recommended policy. Do not infer the approval outcome only from a tool's display name or a broad “read/write” summary. Test the exact operation with non-sensitive data.

See [Security & Privacy](https://docs.odella.ai/workforce/settings-security-privacy.md) for the settings reference and [Approvals](https://docs.odella.ai/workforce/approvals.md) for resolving requests.

## Browser work is a different boundary

Browser access and connected-tool access are not interchangeable. A website button is not automatically subject to the connected-tool approval matrix.

### Public-page and workspace browsing

The workspace browser supports public-page work and local rendering. Its enforced checks include command validation, separation of workspace browser targets, restrictions on unsupported navigation schemes, and checks against direct navigation to common local, private-network, and cloud-metadata addresses.

For local files, browser navigation is restricted to supported file types inside the worker's workspace, with checks against sensitive locations and paths that resolve outside it.

Public-site browsing also has bounded tab, navigation, and action budgets. It stops for detected visible password or one-time-code fields and supported challenge pages. A top-level access denial or rate-limit response restricts further browsing of that origin.

These are safeguards, **not a complete network filter or credential-leak prevention system**. They do not justify promising that every redirect, page subresource, site action, or possible data transfer is blocked.

### Signed-in browser work

Signed-in browsing uses a separate browser environment associated with the authenticated user's account and identity. That user's AI Employees can share that browser context. A run without the required user identity does not receive that user's private browser through the remote browser command boundary.

This distinction matters: a private browser is **user-scoped**, not a separate login vault for every worker. Grant access to sensitive sites deliberately, use restricted website accounts, and review saved sign-ins in [Vault](https://docs.odella.ai/platform/vault.md).

> [!WARNING]
>
> Do not assume that every signed-in website action requires an Odella approval,
> or that every worker has an independent browser login. Browser availability
> does not establish a per-site, per-action approval guarantee.

For website identity selection, saved-session removal, downloads, and safe tests, see [Signed-in browser security](https://docs.odella.ai/platform/browser-security.md).

## Example: draft customer replies without broad access

1. Limit the worker's internal audience to a customer-service team and qualifying role.
2. Give only the manager or designated editors permission to change tools and approvals.
3. Connect a test support account with the minimum provider permissions.
4. Start with **Approval-first**, review saved exceptions, and use **Ask every time** for supported outbound messaging.
5. Ask the worker to read a test record, draft a reply, and attempt a record update. Verify the actual review behavior for each operation.
6. If the work requires a signed-in website, review that website account separately. Do not rely on connected-tool approvals to gate website mutations.

Follow [Create a team-restricted AI Employee](https://docs.odella.ai/platform/secure-employee-tutorial.md) for the full team, role, and access-validation walkthrough.

## Questions for a security review

Ask Odella to confirm requirements that are specific to your deployment or agreement:

* Network destinations and any private-network connectivity requirements.
* Isolation, regional processing, and third-party browser or model dependencies.
* Which exact tool and browser operations require human review.
* Credential scope, saved website sessions, retention, and revocation behavior.
* Available security evidence and audit records beyond the worker Activity Log.

For contractual security measures and assurance resources, see [Security and data protection](https://docs.odella.ai/platform/security.md).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.