MCP integration

The Tharga.MongoDB.Mcp package exposes MongoDB monitoring data and admin actions over MCP (Model Context Protocol). AI agents (Claude, Cursor, etc.) can query collections, inspect monitoring data, and trigger admin actions — without SSH-ing to a prod box for mongosh.

Install + register

dotnet add package Tharga.MongoDB.Mcp

services.AddThargaMcp(mcp =>
{
    mcp.AddMongoDB();
});

app.UseThargaMcp();

The provider is registered on McpScope.System and surfaces only to system-level MCP callers.

Data access levels

Default exposure is metadata only — nothing that returns or modifies actual document data. Opt in for more:

services.AddThargaMcp(mcp =>
{
    mcp.AddMongoDB(o =>
    {
        o.DataAccess = DataAccessLevel.DataRead;       // adds tools/resources that read document data
        // o.DataAccess = DataAccessLevel.DataReadWrite; // also adds tools that modify data (e.g. mongodb.clean)
    });
});

Anything above the configured level is filtered out of tools/list and resources/list, and rejected at tools/call / resources/read with an IsError response.

Resources

URI	Level	Description
`mongodb://collections`	Metadata	Collections with stats, indexes, clean status
`mongodb://clients`	Metadata	Connected remote monitoring agents
`mongodb://monitoring`	DataRead	Recent + slow calls, summaries, error summary, connection-pool state (calls embed filter values, hence the gating)

Tools

Tool	Level	Args
`mongodb.touch`	Metadata	`databaseName`, `collectionName`, optional `configurationName`
`mongodb.rebuild_index`	Metadata	`databaseName`, `collectionName`, optional `configurationName`, `force`
`mongodb.restore_all_indexes`	Metadata	optional `configurationName` / `databaseName` filters
`mongodb.drop_index`	Metadata	drops indexes not declared in code
`mongodb.reset_cache`	Metadata	resets the in-memory monitor cache
`mongodb.clear_call_history`	Metadata	clears recent + slow call history
`mongodb.find_duplicates`	DataRead	`databaseName`, `collectionName`, `indexName`; returns duplicate-key tuples
`mongodb.explain`	DataRead	`callKey` (Guid); returns explain plan + original filter
`mongodb.clean`	DataReadWrite	deletes orphaned / invalid documents
`mongodb.get_document`	DataRead	`id` is auto-detected as Guid → ObjectId → string; returns MongoDB Extended JSON
`mongodb.list_documents`	DataRead	`limit` (default 20, max 200), `skip`, JSON `filter`, JSON `sort`
`mongodb.compare_schema`	DataRead	three-way diff: C# entity properties vs registered type names vs field set observed in sample (default 50 docs, max 500)

Document inspection

mongodb.get_document, mongodb.list_documents, and mongodb.compare_schema let an authorized agent diagnose schema drift and migration fallout via MCP — the same loop typically done via mongosh on a production shell, but through the MCP plumbing you already use for everything else.

All three are DataRead — hidden by default. Opt in via o.DataAccess = DataAccessLevel.DataRead.
Documents are returned as MongoDB Extended JSON — the exact shape stored in MongoDB, never round-tripped through the C# serializer.
compare_schema reflects on the entity type's public properties and compares against the sampled field set. Top-level fields only — nested-document drift is a known limitation.
Per-tenant databases (DatabasePart / per-team DBs) work directly: pass the resolved databaseName from mongodb://collections.
Remote-only collections (Registration.NotInCode) are not yet supported — these tools throw a clear error. Adding remote routing requires extending IRemoteActionDispatcher and the Monitor.Server pipeline; planned as a follow-up.