Site Agent Lets Drupal Editors Review AI-Proposed Website Changes

New Approval Behaviour, MCP Dependencies and More Detailed Change Previews
The Drop Times graphic titled "Site Agent Adds Human Oversight to AI Editing" highlights "Preview changes, control permissions and approve actions in Drupal." An illustrated portrait is labelled "Tyler Brown Cifu Shuster".

Website editors and administrators can use Site Agent to propose changes to Drupal content and configuration through an AI chat interface, with approval and permission controls governing tool execution. The contributed module, developed and maintained by Tyler Brown Cifu Shuster of FivePaths, released version 1.0.4 on 17 November 2026, following version 1.0.3 on 9 November. The latest release changes how certain tool calls seek approval, improves content previews and revision comparisons, and updates dependencies for remote MCP clients.

Site Agent separates AI-requested actions from the permissions and approvals needed to execute them. According to the project's documentation, administrators can configure approval behaviour for tools, including automatically allowed operations and calls requiring human confirmation. Version 1.0.4 was tested against Tool API 1.0.0-beta12; under that integration, only a tool that declares itself destructive is automatically forced to ask for approval on every call. Pre-approved writes that do not declare that flag no longer prompt automatically, so administrators must examine each tool's declared behaviour and configured approval policy rather than assume that every write requires confirmation.

For content editing, the module can show proposed changes and save supported edits as drafts. In version 1.0.4, result cards for a content specification or dry run show the proposed differences, with changed words highlighted and a single Publish control alongside the preview. Comparisons between revisions now identify edits inside Paragraphs entities, not merely changes to fields referencing them; other content writes also refresh the result preview. The project's standard editing workflow keeps publication as a separate action for users with the necessary permissions, but individual previews and approvals should still be tested against the site's content model.

Configuration changes follow a different process. Where configuration is writable, Site Agent can present proposed changes as YAML differences before they are applied. These operations can include changes to content types, fields and other site structures. Configuration-writing tools are unavailable when configuration is locked, although content editing can continue through supported workflows.

The project's approval documentation also distinguishes approval from authorisation. Before an approved operation executes, Site Agent checks the tool arguments and the account's access permissions again. Approval does not grant additional Drupal permissions, and an operation can still be refused if the user no longer has the necessary access or the proposed action has changed.

According to the approval documentation, each tool call leaves an audit record containing its arguments, approval status, decision-maker and execution outcome. Records are stored separately from conversation transcripts, with configurable retention settings. The data and permissions documentation also explains that conversation history, system prompts and tool results are sent to the configured AI provider. Site administrators therefore need to consider which content and configuration information their permitted tools can expose.

The optional Site Agent PHP submodule can run PHP through the interface or a separate console, subject to its restricted permission and configuration. Version 1.0.4 introduces Read with PHP, which the release notes describe as executing code only when a check classifies it as read-only; that tool is pre-approved in the Developer toolset. The release also allows the agent to read and change Paragraphs without PHP, and directs it to use content-model and specification tools before resorting to code. The read-only classification should not be treated as a guarantee of isolation: PHP execution remains a higher-risk capability requiring appropriately trusted accounts and validation.

Site Agent also offers experimental Model Context Protocol (MCP) integration so external AI clients can use the tools of an authorised Drupal account. For its MCP submodule, version 1.0.4 requires mcp/sdk 0.9 and MCP Server 2.0.0-beta7; the release notes identify unauthenticated memory exhaustion in the earlier SDK 0.8 OAuth middleware, tracked as GHSA-f4rf-6ghj-mhr6. Remote approvals now remain in one thread, repeated approval responses resume that thread, and unresolved calls expire. Dry runs and writes not declared destructive no longer cause approval prompts automatically, making client authentication, declared tool capabilities and approval policy important deployment checks.

The module builds on Drupal's AI module, Tool API and Tool Belt. Its project page lists support for Drupal 11.3 and later Drupal 11 releases and for Drupal 12, with PHP 8.3 stated as the baseline requirement; supporting dependencies also need to be compatible with the selected Drupal version. The 1.0.4 release notes specifically identify testing against Tool API 1.0.0-beta12 and updated MCP dependencies. The release also resolves a Content Deployment integration detail by retrieving its SpecDiff service through the registered service identifier, so comparison cards work with content_deployment 1.0.x installations.

Site Agent's maintainers distinguish their implementation from the existing AI Agents module by pointing to its integrated chat interface, approval handling, change previews and tool-call records. These distinctions come from Site Agent's project documentation rather than an independent comparison of the two modules.

Version 1.0.4 also introduces chat bubbles and a narrow-screen drawer fix. These interface changes are secondary to the release's changes to approval prompts, remote-client dependencies and content comparisons. The release notes, which the maintainer says were drafted with an LLM and then reviewed, document intended behaviour rather than independent validation. Teams assessing Site Agent for production should test pre-approved writes, destructive tool declarations, publishing permissions, MCP authentication, data shared with AI providers and PHP capabilities in their own environment. Further details are available in the 1.0.4 release notes and the project documentation.

References

  1. Site Agent 1.0.3 | Drupal.org, Drupal.org (9 November 2026)
  2. Site Agent 1.0.4 | Drupal.org, Drupal.org (17 November 2026)

Disclosure: This content is produced with the assistance of AI.

Note: The vision of this web portal is to help promote news and stories around the Drupal community and promote and celebrate the people and organizations in the community. We strive to create and distribute our content based on these content policy. If you see any omission/variation on this please reach out to us at #thedroptimes channel on Drupal Slack and we will try to address the issue as best we can.

Upcoming Events

Latest Opportunities