The distinction is control
Tools are invoked by the model, so they may have side effects and the model must be able to decide when. Resources are read by the client, usually to put context in front of the model, and must therefore be safe to read at any time. Prompts are chosen by the user — a named workflow that pre-loads instructions.
Side by side
| Tool | Resource | Prompt | |
|---|---|---|---|
| Invoked by | The model | The client | The user |
| May have side effects | Yes | No | No |
| Typical content | An action | Data or documents | Instructions |
| Identified by | A name and a schema | A URI | A name |
| Seen by the user as | A call the model made | Context | A menu item |
When to reach for each
- It changes state, costs money, or is irreversible → tool. Publishing is a tool. So is deleting.
- The model should be able to look at it without deciding to → resource. A list of the caller's sites is a reasonable resource; deleting one never is.
- A multi-step recipe a person picks deliberately → prompt. "Publish this report and gate it to my company domain" is a prompt.
The trap: read-only tools
The common mistake is a read-only tool where a resource belongs — get_site as a tool
rather than the site as a resource. It works, and it costs you: the model has to decide to call
it, so it often does not, and when it does it burns a turn. A resource can just be there.
The opposite mistake is worse. Anything with a side effect exposed as a resource can be triggered by a client that is merely gathering context. Read-safety is the contract.
A publish-shaped server, sorted
Applying the rule here: the twelve verbs are tools, because each changes something or reads something the model specifically asked for. Documents that describe the deployment — the endpoint list, the defaults, the page index at /llms.txt — are resource-shaped, and are published as plain HTTP so any client can read them whether or not it implements the resource primitive.
Not every client implements all three
Tool support is universal; resources and prompts are patchier, and the gap moves. The practical consequence: if a capability is essential, expose it as a tool, because that is the primitive you can rely on. Use the other two to make life nicer for clients that have them.
Can the same thing be both a tool and a resource?
Yes, and for read-only data that is often pragmatic while client support is uneven. Keep the behaviour identical so the model cannot tell which door it came through.
Do resources have to be text?
No — a resource has a URI and a MIME type. Text and JSON dominate because that is what a model can act on.
Which does this deployment expose?
Twelve tools, plus discovery and description documents over plain HTTP. The verbs are where the value is, and they are the primitive every client can call.
Are prompts worth implementing?
Only if the workflow is genuinely multi-step and users repeat it. Otherwise it is a menu item duplicating what the model would have done anyway.