Documentation

Source Connections and Tool Policy

View as Markdown

Source connections and tool policy define what an Agent can use during a run. This is where you turn a repeatable AI process into a controlled process over approved sources.

Source Connections

Source connections identify the content or systems available to the Agent.

Examples include:

  • selected Clear Ideas Sites
  • folders or content scopes within a Site
  • Clear Ideas MCP context
  • approved external MCP document connectors
  • approved web or API tools

Keep source scope as narrow as the task allows.

Create and authenticate reusable saved connections in Agent Connections, then attach only the needed connections to each Agent.

Required Connections

Mark a connection as required when the Agent should not run unless that source is healthy and available. Required connections are useful for scheduled reports, external deliverables, and processes where missing context would make the output misleading.

Read and Write Modes

Use read-only mode for most source connections. Enable write actions only when the Agent is expected to create folders, save files, update records, or send results to an external system.

Write-capable Agents should have clearer review gates, narrower credentials, and more explicit evidence settings.

Allowed Tools

Allowed tools limit what the Agent can do through each connection.

Examples:

  • list folders
  • search content
  • read document text or summaries
  • save generated files
  • call a webhook
  • create a folder
  • use web search

Disable tools that are not needed for the Agent's purpose.

Source Instructions

Source instructions tell the Agent how to use each source. Good instructions explain:

  • what the source contains
  • when to use it
  • what not to use it for
  • how to handle missing or conflicting information
  • whether source citations are required

Evidence Settings

Evidence settings determine how source usage, generated outputs, files, and tool calls are recorded. For sensitive or external-facing work, prefer settings that preserve source IDs, source titles, retrieval metadata, generated file hashes, and final output references.

See Governed AI Records.