Documentation

Controlled Tools and Egress

View as Markdown

Controlled tools and egress settings limit what an Agent can call, where it can send information, and what evidence is recorded. Use these controls when Agent runs involve sensitive documents, external stakeholders, or regulated decisions.

Tool Control

Tool control starts with an allowlist. An Agent should only have access to the tools it needs for the task.

Common tools include:

  • Site search
  • content read or summary tools
  • generated file creation
  • save file actions
  • web search
  • webhook calls
  • external MCP tools

Egress Control

Egress controls limit outbound activity. They matter when an Agent can send data outside Clear Ideas, call external systems, or use tools that retrieve public web content.

Review egress settings before enabling:

  • outgoing webhooks
  • write-capable external MCP tools
  • web research
  • file save operations to external systems
  • integrations with customer, CRM, CMS, or ticketing systems

Web Access

Web access should be enabled only when current public information is required. For document-grounded work, prefer approved Site content and require citations to the governed sources.

When web access is allowed, tell the Agent which claims should come from approved content and which claims may use current public sources.

Generated Files

Generated spreadsheets, documents, presentations, and images are governed outputs. They should remain attached to the chat or Agent run that produced them, with metadata and hashes available for evidence export when required.

See File Generation in Chat and Agents.

Review Gates

Use review gates when an Agent can produce external-facing deliverables, call write tools, or send outbound webhook payloads.

Review gates are especially important for:

  • board or investor materials
  • legal or compliance reviews
  • customer communications
  • financial reporting
  • any Agent with write-capable tools

Evidence

For audit-ready Agent work, preserve the source context, tool calls, generated files, final output, and policy settings needed to understand what happened.

See Governed AI Records.