A read-only MCP setup lets the assistant request information without offering a tool that edits it. You can see what this looks like in Blender by exposing one tool that lists scene details. The same approach works for a course catalogue that the assistant can search without changing any entries.
Step-by-step setup
Limit Blender MCP to a scene lookup
What you need The working local Blender connection from the companion guide and a Codex setup that supports a tool allowlist.
Confirm which server you registered
Use the
blenderserver created with the CLI instructions. A plugin-provided server has a different configuration section, so follow the host documentation for that case.Allow only the scene tool
Open the existing Codex
config.toml. Under its existing[mcp_servers.blender]section, add the line below. Keep the server’s command and arguments in that same section. Start a new Codex session.Check what the agent can call
Use the MCP tool list for that server. It should expose
get_scene_infoand leaveexecute_blender_codeunavailable through this connection. Close any second Blender connection that would still expose edit tools.Try it on a fresh scene
In Blender, open a General practice scene and rename Cube to PracticeCube. Ask the assistant to list the object names through
get_scene_info, then compare the response to the Outliner.
Add this line under the existing server entry
enabled_tools = ["get_scene_info"]
How to tell it worked
The lookup returns the known object list and the Blender connection exposes only the scene lookup. If another tool can still edit Blender, check which server entry it belongs to before continuing.
This setting limits what Codex can call through this entry. It does not change the permissions of the Blender add-on itself or another connected app.
Official setup references
Define the smallest useful tool
Use a tool that accepts a known course ID or a short search phrase and returns approved topics with their source references. Give it a description that explains when to call it and what to do if there is no match. Do not give the tool access to student records or administrative functions for this exercise.
Limit what the server can reach
Choose credentials and an account that can only read the practice catalogue. Review which capabilities the host discovers when it connects. Where approvals or permissions are available in the host, use them to control how a tool is called. The actual data access also needs to be limited on the server side.
Handle uncertain and irrelevant results
Test an exact name, a similar course, an unknown course and a question the catalogue cannot answer. Use a clear “not found” result rather than a guessed match. Text returned by a resource is source material. It should not override the workflow’s instructions or access rules.
Record enough to review a run
Log the tool name, course identifier, result status and source reference without storing unrelated content. Check whether an answer based on a match points back to the same source. Document how staff can disconnect the tool or update the approved catalogue later.
Try it yourself
Create three sample courses and one intentionally ambiguous query. Confirm that an unknown query gives a useful fallback and that a known query references only the approved document it found.
Common questions
Do I need access to a live database?
No. A small sample file or test catalogue is enough to practise modelling the tool and checking the responses.
What does read-only mean here?
The server can retrieve approved catalogue entries. The credentials and the tool implementation should both prevent changes to those entries.
