On 16 November 2026 the GLPI team released version 2.2.0 of the official GLPI AI plugin. It looks like a minor release, but three changes matter to administrators: a customizable system prompt, less user data sent to external LLM providers, and a medium-severity security fix. If you run the plugin, schedule the update soon.
What is new in 2.2.0
Custom system prompt
The assistant's system prompt can now be edited directly in the plugin settings. That lets you describe your environment, terminology, service catalog or preferred answer style. A changed prompt applies only to new chat threads, and existing threads keep the old one. The release notes also warn that some frontier models enforce safety or topical constraints based on system instructions, so it pays to write the prompt with a clearly defined domain context.
Less metadata sent to AI providers
According to the release notes, the plugin now sends less user account metadata to external LLM providers. For organizations working on data minimization that is a step in the right direction. The release does not say which fields were dropped, though, so it does not replace your own data flow assessment.
Performance and a security fix
- Ticket context reaches the model more efficiently, which means fewer MCP tool calls and faster responses.
- GLPI MCP tool outputs were reformatted to support more reliable model reasoning.
- A medium-severity fix: a missing rights check on the Writing Action.
A reminder: custom and self-hosted models
Version 2.1.0 already added support for self-hosted and third-party endpoints compatible with the OpenAI API. The release notes name LiteLLM, vLLM and LocalAI as examples. For the public sector and regulated entities this is an important option: with a self-hosted endpoint, ticket data does not have to leave your infrastructure. The notes do warn that model capabilities, stability and advanced features such as vector embeddings may vary by backend.
What the plugin does today and where to be careful
Per the documentation (as of 22 November 2026), the plugin supports GLPI 11.0.x and 12.0.x and requires a GLPI Network BASIC subscription or higher. It offers summaries, translation, a writing assistant, category, impact and urgency suggestions, related knowledge base articles and tickets, and a chat that can read, search and update GLPI data through MCP tools. In practice, these are the points to check:
- If you enable automatic classification (category, impact, urgency), the documentation says it runs whenever an ITIL item is created – a ticket, problem or change – and erases any predefined values. If templates or business rules already set the category, test how the two interact before turning it on.
- Approvals in chat: data-altering actions (anything that is not a read) and calls to external MCP servers require user approval. Make sure users know what they are approving.
- The debug log writes requests, responses and MCP calls to glpiai-debug.log. That includes ticket content, so enable it only temporarily and handle the file as sensitive data.
- Outbound traffic: the updateModelsDevCatalog automatic action needs outbound HTTPS to models.dev. On closed networks, disable it or allow it at the firewall.
- Access: the setup wizard offers "just me", "all technicians" or "nobody". A narrow group is enough for a pilot.
Suggested approach
Roll out 2.2.0 in a test environment first; since it includes a security fix, don't delay the production update. Then draft a system prompt that describes your services and communication rules. For entities under NIS2 it makes sense to document which model provider you use, what data it receives and who has access to the AI features. Where you can, consider a self-hosted endpoint.