MongoDB MCP — Use Cases, Install & Live Demo

Why use it

Key features

Works against MongoDB Atlas, Enterprise, or Community — same tools
CRUD + aggregation pipelines with full query/projection/sort support
Atlas admin surface: projects, clusters, DB users, IP allowlist
Read-only by default; writes opt-in via --read-only flag
Schema inference: infer a collection's shape by sampling

Live Demo

What it looks like in practice

mongodb.replay ▶ ready

0/0

Install

Pick your client

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "mongodb": {
      "command": "npx",
      "args": [
        "-y",
        "mongodb-mcp-server"
      ]
    }
  }
}

Open Claude Desktop → Settings → Developer → Edit Config. Restart after saving.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "mongodb": {
      "command": "npx",
      "args": [
        "-y",
        "mongodb-mcp-server"
      ]
    }
  }
}

Cursor uses the same mcpServers schema as Claude Desktop. Project config wins over global.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "mongodb": {
      "command": "npx",
      "args": [
        "-y",
        "mongodb-mcp-server"
      ]
    }
  }
}

Click the MCP Servers icon in the Cline sidebar, then "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "mongodb": {
      "command": "npx",
      "args": [
        "-y",
        "mongodb-mcp-server"
      ]
    }
  }
}

Same shape as Claude Desktop. Restart Windsurf to pick up changes.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "mongodb",
      "command": "npx",
      "args": [
        "-y",
        "mongodb-mcp-server"
      ]
    }
  ]
}

Continue uses an array of server objects rather than a map.

~/.config/zed/settings.json

{
  "context_servers": {
    "mongodb": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "mongodb-mcp-server"
        ]
      }
    }
  }
}

Add to context_servers. Zed hot-reloads on save.

claude mcp add mongodb -- npx -y mongodb-mcp-server

One-liner. Verify with claude mcp list. Remove with claude mcp remove.

Use Cases

Real-world ways to use MongoDB

Answer business questions with Mongo aggregation pipelines

👤 PMs and analysts on a Mongo-backed product ⏱ ~15 min beginner

When to use: You need counts, funnels, or top-N lists and don't want to learn $group/$lookup syntax.

Prerequisites

Read-only connection string — Atlas: create a DB user with readAnyDatabase. Self-hosted: user with read role on relevant DBs.

Flow

Discover collections

List databases, then for app_prod list all collections and their approximate document counts.✓ Copied

→ Collection catalog
Sample and infer schema

Sample 20 docs from users and orders. Describe the fields and types you see.✓ Copied

→ Per-collection schema description
Run the actual aggregation

How many orders were placed per country in the last 30 days? Sort desc and limit 20.✓ Copied

→ Results table with the pipeline used

Outcome: Business answers with the exact pipeline preserved for re-running.

Pitfalls

Aggregations without indexes can scan huge collections — Always check .explain() first and make sure a supporting index exists; otherwise add $match narrowly on an indexed field at the top of the pipeline

Combine with: notion

Infer and document a messy collection's actual schema

👤 New engineer onboarding to an undocumented Mongo ⏱ ~25 min intermediate

When to use: The collection was 'schemaless' for 3 years and nobody knows which fields actually exist.

Flow

Sample broadly

Sample 500 docs from events. For each top-level field, report presence %, type(s), and a sample value.✓ Copied

→ Field-by-field presence/type matrix
Find schema drift

Which fields have multiple types across docs? Group by (field, type) and count.✓ Copied

→ List of polymorphic fields
Produce a TypeScript type or JSON schema

Generate a TypeScript interface for the 'stable' fields (≥95% presence, single type). Mark the rest as optional or unknown.✓ Copied

→ Usable type definition

Outcome: A documented schema with known quirks — the basis for migrations or a validator.

Pitfalls

500 docs may miss rare but important variants — Sample by time bucket (one per month) to catch legacy shapes

Combine with: filesystem

Audit your Atlas projects for security and cost

👤 DevOps/platform teams on Atlas ⏱ ~20 min intermediate

When to use: Quarterly: check which clusters are oversized, which have wide IP allowlists, who has access.

Prerequisites

Atlas API public+private key — cloud.mongodb.com → Organization Access → API keys; project-scoped

Flow

List projects + clusters

List every project and within each, every cluster with its tier, region, and backup status.✓ Copied

→ Full inventory
Flag risky access

For each project, dump the IP access list. Flag any entry of 0.0.0.0/0 with the project name.✓ Copied

→ Risky-access report
Suggest rightsizing

Any cluster on M30+ with fewer than 10GB used? Recommend downgrades.✓ Copied

→ Cost-savings list

Outcome: A short remediation list for your security + finance folks.

Pitfalls

API key scope too narrow to see every project — Use an Organization-level key in read-only mode rather than a project-level one

Combine with: notion

Propose and execute a one-off data cleanup safely

👤 Backend engineer fixing a data bug ⏱ ~30 min advanced

When to use: A bug caused bad writes; you need to fix ~10k docs but must not nuke the wrong ones.

Prerequisites

Writable user (scoped to the target DB only) — Atlas: role with readWrite on that one DB, nothing else
--read-only OFF explicitly for this session — Start the MCP without --read-only

Flow

Scope the fix with a count

Count docs in users where status='active' AND last_login IS NULL AND created_at < 2024-01-01. Don't modify anything.✓ Copied

→ Expected-affected count, e.g. 9,873
Dry-run the update

Show the updateMany pipeline you would run (filter + $set), and show 5 sample docs that would be changed. Do NOT execute.✓ Copied

→ Filter + set preview
Execute with a limit and verify

Run the update. Then re-run the original count — it should be 0. Report matchedCount and modifiedCount.✓ Copied

→ Counts match expectation; verify query returns 0

Outcome: A clean, auditable fix with counts before and after.

Pitfalls

updateMany with a bad filter nukes the whole collection — Always run the filter as countDocuments first; if the count is surprising, stop and investigate
No backup of the affected slice — Copy the matching docs to a <collection>_backup_<date> collection before updating

Combine with: filesystem

Suggest missing indexes from slow-query patterns

👤 Backend engineers hunting perf issues ⏱ ~25 min advanced

When to use: Your app is slow on Mongo queries; you want a targeted index plan, not a shotgun.

Flow

Check existing indexes

For orders and users, list every index with its keys and size on disk.✓ Copied

→ Index catalog
Profile a specific query

Run .explain('executionStats') on this query [paste]. Report totalDocsExamined vs nReturned and the winning plan stage.✓ Copied

→ Explain output
Propose the smallest new index

Given that plan, propose exactly one index that would convert this to IXSCAN. Justify field order.✓ Copied

→ Concrete createIndex command with rationale

Outcome: A single, justifiable index recommendation per slow query — not a wall of them.

Pitfalls

Over-indexing kills write throughput and balloons storage — Only add an index that serves >1 high-traffic query; compound indexes should reflect ESR (Equality, Sort, Range) order

Combinations

Pair with other MCPs for X10 leverage

mongodb + notion

Aggregate, then post a shareable report

Compute MAU per plan tier for the last 6 months and create a Notion page in 'Growth / Monthly' with the results as a table.✓ Copied

mongodb + filesystem

Backup a collection slice as JSONL before a cleanup

Find all docs in users matching <filter>, save to /backups/users-cleanup-<date>.jsonl, then delete them.✓ Copied

mongodb + postgres

Cross-DB reconciliation when migrating off Mongo

For every user_id in Mongo users, check whether a corresponding row exists in Postgres users. Report mismatches.✓ Copied

Tools

What this MCP exposes

Tool	Inputs	When to call	Cost
list_databases		Start of any exploration session	free
list_collections	database: str	Inventory a database	free
find	database, collection, filter?, projection?, sort?, limit?	Read documents — the main read workhorse	free
aggregate	database, collection, pipeline: stage[]	Grouping, joining, analytics	free
count	database, collection, filter?	Always before destructive writes — confirm scope	free
insert_one / insert_many	database, collection, document(s)	Requires --read-only off	writes
update_one / update_many	database, collection, filter, update	Always preview filter with count first	writes
delete_one / delete_many	database, collection, filter	Dangerous — require explicit user confirmation	writes
list_indexes	database, collection	Perf analysis before suggesting new indexes	free
atlas_list_projects / atlas_list_clusters		Atlas control-plane audits	free

Cost & Limits

What this costs to run

API quota: Driver: bounded by cluster connection limit. Atlas API: 100 req/min per key.
Tokens per call: Find/aggregate: scales with result size; use projections and limits.
Monetary: Free against your existing cluster. Atlas has a free M0 tier for testing.
Tip: Always project only the fields you need; unbounded finds return large docs that chew context and egress.

Security

Permissions, secrets, blast radius

Minimum scopes: readAnyDatabase (read-only) or read on specific DBs

Credential storage: MDB_MCP_CONNECTION_STRING for driver; MDB_MCP_API_CLIENT_ID + MDB_MCP_API_CLIENT_SECRET for Atlas

Data egress: Driver connects to your cluster; Atlas API to cloud.mongodb.com only

Never grant: dbAdminAnyDatabase userAdminAnyDatabase root

Run with --read-only for exploratory sessions; only disable for explicit fix work.
Never point at prod primary for heavy aggregations; use a hidden/analytics node or Atlas Online Archive.

Troubleshooting

Common errors and fixes

MongoServerError: Authentication failed

Connection string user/password wrong or the user lacks auth DB. Add ?authSource=admin for Atlas.

Verify: mongosh '$MDB_MCP_CONNECTION_STRING' --eval 'db.runCommand({ping:1})'

MongoNetworkError: ETIMEDOUT

IP not in Atlas allowlist. Add your current IP in Atlas → Network Access.

Verify: curl ifconfig.me then compare

not authorized on admin to execute command listDatabases

Role too narrow. Grant clusterMonitor or scope tools to a specific DB via listCollections instead.

Write rejected / running in read-only mode

Restart the MCP without --read-only; only do this for the specific fix session.

Alternatives

MongoDB vs others

Alternative	When to use it instead	Tradeoff
Postgres MCP	You're on Postgres, or considering migrating from Mongo	Relational — document-style flexibility gone
DBHub	You need one MCP for Mongo + several SQL DBs	Shallower Mongo feature coverage than the official server

More

Resources

📖 Read the official README on GitHub

🐙 Browse open issues

🔍 Browse all 400+ MCP servers and Skills