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

# Boundary drift

> Detect when your agent uses undeclared tools or touches unknown resources in production.

You declared what your agent should do. Now you want to know when production behavior diverges — a new MCP tool appears, an agent starts calling `deploy.release` without a declaration, or a resource type you've never seen shows up in telemetry.

Ratri compares runtime tool calls against your declared capabilities and surfaces drift warnings.

## Enable drift warnings

Configure boundary warnings in your SDK config:

<CodeGroup>
  ```ts TypeScript theme={null}
  const ratri = new Ratri({
    agent: { key: "my-agent", name: "My Agent" },
    boundary: {
      warnOnUndeclaredTools: true,
      warnOnUnknownResourceTypes: true,
      autoInferFromToolNames: true,
    },
  });
  ```

  ```python Python theme={null}
  ratri = Ratri.create({
      "agent": {"key": "my-agent", "name": "My Agent"},
      "boundary": {
          "warn_on_undeclared_tools": True,
          "warn_on_unknown_resource_types": True,
          "auto_infer_from_tool_names": True,
      },
  })
  ```
</CodeGroup>

| Setting                      | What it detects                                                                           |
| ---------------------------- | ----------------------------------------------------------------------------------------- |
| `warnOnUndeclaredTools`      | Tool calls for tools not in your declared capabilities                                    |
| `warnOnUnknownResourceTypes` | Resource types not seen in any capability declaration                                     |
| `autoInferFromToolNames`     | Infer action/resource from tool names (helps match undeclared tools to expected patterns) |

## What drift looks like

### Undeclared tool

Your agent calls `vault.read_secret` but you only declared `search` and `github.merge_pr`:

* Ratri emits the tool call event normally
* A boundary warning is attached: tool `vault.read_secret` is not in declared capabilities
* In the dashboard, the boundary map highlights the gap

### Unknown resource type

Your agent touches `shell_command` but your capabilities only list `code_repository` and `work_item`:

* A warning flags the unknown resource type
* Guardrails may treat it as higher risk if templates match on resource type

### What you'll see

Boundary drift warnings in the dashboard boundary map and in event validation output from `doctor`.

## Remediation workflow

1. **Monitor** — run in monitor mode with drift warnings enabled
2. **Review** — check which undeclared tools appear in production telemetry
3. **Declare** — add missing capabilities via config or `capabilities declare`
4. **Enforce** — enable guardrail templates and switch to Enforce mode

<CodeGroup>
  ```bash TypeScript theme={null}
  npx ratri capabilities declare
  npx ratri guardrails enable prod-secrets
  ```

  ```bash Python theme={null}
  ratri capabilities declare
  ratri guardrails enable prod-secrets
  ```
</CodeGroup>

## MCP auto-discovery

The MCP proxy defines tools automatically when it receives `tools/list` from the upstream server. This reduces drift for MCP-hosted agents — but review auto-discovered tools before enabling Enforce mode.

See [MCP proxy](/mcp/proxy).

## Next steps

<CardGroup cols={2}>
  <Card title="Declare capabilities" icon="list-check" href="/boundaries/declare-capabilities">
    Add missing tool declarations.
  </Card>

  <Card title="Boundary reports" icon="file-chart-column" href="/production/boundary-reports">
    Generate compliance reports over a time window.
  </Card>
</CardGroup>