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

# Brane vs Prompt Guardrails

> Brane differs from prompt guardrails by enforcing policy at the AI agent action boundary before tools, model calls, database queries, and MCP tools execute.

Brane is different from prompt guardrails because Brane enforces policy at runtime when an agent attempts an action. Prompt guardrails influence model behavior. Brane controls whether the action is allowed to execute.

Both can be useful, but they protect different boundaries.

## Comparison

| Control layer        | What it checks                  | When it runs                     | What it prevents                  |
| -------------------- | ------------------------------- | -------------------------------- | --------------------------------- |
| Prompt guardrail     | Instructions and model behavior | Before generation                | Some unsafe reasoning or output   |
| Output filter        | Text returned by the model      | After generation                 | Some unsafe text                  |
| Brane runtime policy | Tool calls and capabilities     | Before or after action execution | Unsafe actions and capability use |

## Why Runtime Policy Matters

If an agent sends an email, deletes a user, issues a refund, queries a database, calls an MCP tool, or writes memory, the consequence happens at the action boundary.

Brane evaluates policy at that boundary:

```text theme={null}
Capability + AgentAction + PolicyContext -> Policy -> Decision
```

That gives developers a concrete enforcement point.

## Use Brane With Existing Safety Layers

Brane does not replace:

* Prompt engineering
* Model safety systems
* IAM
* Database permissions
* Network controls
* Secrets management
* Application authorization

Brane complements those layers by governing agent-specific action patterns.

## Related Pages

* [AI Agent Guardrails](/guides/agent-guardrails)
* [Policy-as-Code For AI Agents](/guides/policy-as-code-for-ai-agents)
* [Mental Model](/concepts/mental-model)
* [Current Status](/current-status)