microsoft-graph-querying

Querying Microsoft Entra with the Graph Enterprise MCP

The Microsoft Graph MCP Server for Enterprise is not a raw Graph API passthrough. It is built around a Retrieval-Augmented-Generation (RAG) workflow: you describe what you want, the server hands you vetted candidate API calls from a curated catalog, and you execute the one that fits. Following this loop is the difference between correct answers and made-up endpoints.

The core rule

Always start with microsoft_graph_suggest_queries. Do not invent raw Graph endpoints.

Microsoft Graph is enormous and its URL/$filter/$select syntax is full of sharp edges (advanced query parameters, ConsistencyLevel headers, OData casts). The suggest_queries tool exists precisely so the model doesn't have to guess. Hand-writing a Graph URL and passing it to microsoft_graph_get is an error pattern — even when you "know" the endpoint, route through suggest_queries so you get the catalog's correct, tested form.

The three tools

Tool	Role in the loop
microsoft_graph_suggest_queries	RAG search over a curated catalog of Graph API examples. Input: a natural-language intent. Output: candidate Graph API calls (endpoint, parameters, description) ranked by relevance.
microsoft_graph_get	Executes a read-only Graph GET. Input: a Graph API call (ideally one returned by suggest_queries). Honors the caller's roles, granted scopes, and Graph throttling.
microsoft_graph_list_properties	Returns the schema for a Graph entity — its properties and relationships. Use it when you need to know what's selectable before constructing or refining a call.

The query loop

1. suggest_queries(intent)   → candidate Graph calls
2. (model)  select the candidate that best matches the question
3. get(selected candidate)   → data
4. (optional) list_properties(entity)  → schema, to refine $select or pick the right entity
5. (model)  translate the JSON result into a plain-language answer

Steps 2 and 5 are your job — suggest_queries proposes, the model disposes. When several candidates look plausible, prefer the one whose description most precisely matches the user's intent and whose parameters you can fill confidently. If a candidate needs a property you're unsure exists, call list_properties first.

Worked examples

"How many users are in the tenant?"

1. microsoft_graph_suggest_queries("count of all users in the tenant") → returns candidates such as a call against the users collection with a count.
2. Pick the candidate that returns a count rather than a full user list (cheaper, directly answers the question).
3. microsoft_graph_get(<that candidate>).
4. Answer: "Contoso has 412 licensed and unlicensed user accounts in Entra ID."

"Which users don't have MFA registered?"

1. microsoft_graph_suggest_queries("users who have not registered for multi-factor authentication") → expect candidates around authentication-methods registration reporting (admin reporting surface).
2. The right candidate is typically a user-registration-details report filtered to users where MFA is not registered — not a per-user method enumeration.
3. microsoft_graph_get(<that candidate>).
4. If you need to also show each user's department or job title, call microsoft_graph_list_properties("user") to confirm those properties, then re-suggest/refine.
5. Answer with a named list and a count: "9 of 412 accounts have no MFA method registered — including 2 with admin roles, which is the priority to fix."

"List the guest (external) users."

1. microsoft_graph_suggest_queries("external guest users in the directory") → candidates filtering the users collection on guest user type.
2. microsoft_graph_get(<that candidate>).
3. Answer with the guest accounts, their invitation/external state, and a count. Flag guests with no recent sign-in as cleanup candidates.

"Find inactive accounts that still hold Copilot licenses."

This needs two reads joined by the model:

1. microsoft_graph_suggest_queries("user accounts with no recent sign-in activity") → an inactive-users / sign-in-activity reporting candidate. Run it with microsoft_graph_get.
2. microsoft_graph_suggest_queries("users assigned a Microsoft 365 Copilot license") → a licensed-users candidate. Run it with microsoft_graph_get.
3. The model intersects the two result sets: accounts that appear in both are inactive and Copilot-licensed — wasted spend.
4. Answer with the overlap list and the reclaimable license count.

The example endpoints above are illustrative of what suggest_queries typically returns — don't hard-code them. Always run suggest_queries for the live, tenant-appropriate candidate.

Using list_properties

Call microsoft_graph_list_properties when you need the schema of an entity — for example to know whether user exposes signInActivity, assignedLicenses, or createdDateTime, or what relationships group has. Use it to:

• choose the right $select properties so a get returns exactly what the question needs;
• confirm an entity actually has the field you're about to filter on;
• decide which entity to query when a question could map to several (user vs servicePrincipal vs device).

Constraints to respect

• Read-only. Everything here reports. There is no way to create, edit, or delete directory objects through this server. If the user asks for a change, say so and point them to a write-capable tool (e.g. the cipp plugin).
• RBAC-honoring. Results are scoped to the signed-in caller's Entra roles and the app's granted scopes. If a query returns less than expected, it may be a permissions boundary, not missing data — and per-tenant admin consent must already be in place (see the microsoft-graph-connection skill).
• Rate limit. 100 calls/min/user plus Graph's own throttling. Don't fan out speculative get calls — suggest_queries exists so one good call beats ten guesses.
• Preview. The example catalog and behavior may change; verify anything material in the Entra admin center.

Presenting results

Raw Graph JSON is not an answer. Always translate: lead with the direct answer (a count, a yes/no, a named list), then supporting detail, then — if it's an audit-style question — a recommendation. Non-technical readers should never see GUIDs or JSON unless they ask.