Model Context Protocol servers that connect AI assistants to real services and data. Browse by domain.
Ranked by Discovery Score — our 0-100 composite of adoption, maintenance, community, and license health. How scoring works.
By Erik Loyd, SaaS CEO and former COO/CFO of an AWS Premier Partner.
FastMCP is the fastest way to build MCP servers in Python. Decorate a function and it becomes a tool your AI assistant can call, with schema generation, validation, auth, and transport handled for you. Apache 2.0 and free. Version 4 turned it from a server library into a full framework with three pillars: servers that wrap your functions, clients that connect to any server, and Apps that render interactive UIs directly in the conversation. Setup is one uv add and a few lines of code. Written a Flask route? You can write an MCP server. FastMCP 1.0 was folded into the official MCP Python SDK back in 2024, and the standalone project has kept moving well past it, so read the upgrade guides before assuming existing code carries forward. There is an official TypeScript counterpart from the same team if Python is not where you live. Free at every size, and the library is complete rather than a demo that funnels you toward a paid tier. Prefect sells Prefect Horizon separately: an enterprise MCP gateway that deploys servers from GitHub with branch previews, keeps a private registry of every MCP in the company, and adds SSO, tool-level RBAC, and audit logs. Free for personal projects, enterprise pricing on request. Solo builders and small teams will not need it. Anyone running many servers across many teams is buying governance, not the framework. The catch is churn. Version 4 is in beta while 3.4.x is still shipping patches, and the docs carry separate upgrade guides from FastMCP 3, FastMCP 2, and two different MCP SDK generations. That is the cost of being the default in a protocol that is still moving. Pin your version and read the release notes before you bump.
The framework is free and complete. Prefect Horizon is a separate enterprise gateway, priced on request, and you only need it once MCP governance becomes somebody's job.
**Free tier:** The framework is Apache 2.0 and complete. Servers, clients, Apps, auth, and transports all ship in the open source package. No key, no seat count, no feature gate. **Self-hosted:** This is the normal path and it is cheap. A FastMCP server is a Python process, so it runs anywhere Python runs: a container, a VM, a laptop for local tools. Your costs are whatever hosts the process plus whatever API the server talks to on your behalf. The operational work is version discipline more than infrastructure, because the project moves fast and there are upgrade guides from three prior generations. **Paid:** Prefect Horizon is the commercial product, and it is a separate gateway rather than hosted FastMCP. It deploys servers from GitHub with branch previews and rollback, gives you a private registry of every MCP server in the org, and enforces SSO and tool-level RBAC with audit logging across clients like Claude, Cursor, and VS Code. Free for personal projects, enterprise tiers priced through sales. You are buying governance across many servers, not the ability to build one.
Connects your AI coding agent directly to Figma designs. Instead of screenshots or manual descriptions, the model reads your actual design files, including layout structure, styles, and component hierarchies. It bridges the gap between design and implementation. You get read access to Figma files, frames, and components with structured context the model can act on. Requires a Figma access token. The MCP server is free, Figma itself requires a paid plan for most team usage. Maintained by the community (GLips). This is the go-to Figma integration for AI coding workflows. If you are translating Figma designs into code, this saves serious time. Worth installing.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Framework for building full MCP applications with interactive React widgets. Works across ChatGPT, Claude, and other MCP-compatible clients. Think of it as a starter kit for MCP app development. Scaffold a new project with one npx command. Includes 15+ templates, a web debugger, and multi-language support. TypeScript and Python both work. More opinionated than FastMCP, more focused on end-user apps than raw server tooling. Maintained by the Manufact team. Free and MIT licensed. Optional paid cloud hosting available but not required.
## Free Tier The MCP server itself is completely free and open source. Install it, configure your MCP client, and use it without any cost. ## Self-Hosted Runs locally or on your own infrastructure. No external dependencies beyond the service it connects to. ## Paid No paid tier. The MCP server is free. The underlying service may have its own pricing.
The Git server lets an AI assistant work a local repository directly: status, staged and unstaged diffs, diffs against other branches, log with date filters, show, branch listing and creation, checkout, add, reset and commit.
Run it with uvx mcp-server-git against any local repo. No GitHub account, no API keys, no costs. It has no push, pull or other remote operations.
The MCP steering group maintains it, and its README says it is in early development with tools subject to change. Install it when your client has no Git tooling of its own. Coding agents that already run git through a shell get little from it.
Free with no service behind it. Early development, and local only: no push, pull or remote operations.
**Free tier:** Free and open source (the project is moving from MIT to Apache 2.0). No underlying service. **Self-hosted:** Runs locally with `uvx`, `pip` or Docker against repositories on your machine. **Paid:** None.
The Everything server is a test rig, not a tool. It exercises the MCP features a client has to handle: tools, prompts, resources, sampling, elicitation, logging, progress notifications and argument completions, so client builders can check their work against something known.
One npx command runs it over stdio, SSE or streamable HTTP. There is no service behind it and nothing to pay for.
The MCP steering group maintains it as a reference implementation. Building or debugging an MCP client? Install it. Everyone else can skip it, and nobody should run it where the environment holds secrets: its get-env tool returns every environment variable to the model.
Nothing to pay for. Its get-env tool hands every environment variable to the model, so keep secrets out of its environment.
**Free tier:** Free and open source (the project is moving from MIT to Apache 2.0). No underlying service. **Self-hosted:** Runs locally through `npx` or as an installed npm package, over stdio, SSE or streamable HTTP. **Paid:** None.
Fetch hands your AI assistant a URL reader. The model asks for a page, and the server returns it converted to markdown, in chunks the model can page through when the content runs long.
Setup is one uvx mcp-server-fetch entry in your config, or a Docker image. It obeys robots.txt for model-initiated requests unless you pass --ignore-robots-txt, and it takes a custom user agent and a proxy. No API keys, no service fees.
The MCP steering group maintains it as a reference implementation, and its README carries a caution worth reading twice: it can reach local and internal IP addresses. Fine on a laptop. On any network with internal services, sandbox it first.
No service fees. The real cost is risk: it can fetch internal IP addresses, so sandbox it on any network with private services.
**Free tier:** Free and open source (the project is moving from MIT to Apache 2.0). No API keys and no paid service behind it. **Self-hosted:** Runs locally with `uvx`, `pip` or Docker. Installing Node.js is optional and switches it to a more robust HTML simplifier. **Paid:** None. The only cost is exposure: the server can request local and internal IP addresses, so treat it as a network-capable process.
Filesystem lets an AI assistant read and change files, but only inside directories you allow. It covers reading text and media, writing, pattern-based edits with a dry-run preview, creating and listing directories, moving files, search and directory trees.
Set the allowed paths as arguments, or let a client that supports MCP Roots update them at runtime. It starts with npx, or in Docker with mounts you can mark read-only. Local and free.
The MCP steering group maintains it as a reference implementation, not a production product. For local work it is the obvious first install. Point it at a narrow folder, because write_file overwrites existing files.
Nothing to pay. Scope the allowed directories tightly, since write_file overwrites existing files.
**Free tier:** Free and open source (the project is moving from MIT to Apache 2.0). No underlying service. **Self-hosted:** Runs locally through `npx` or Docker. Access is limited to directories passed as arguments or supplied by the client through MCP Roots; the server refuses to start with neither. Docker mounts can be made read-only. **Paid:** None.
Memory gives your AI assistant a notebook that survives between chats. It stores a small knowledge graph of entities, the relations between them and facts attached to each, then lets the model search and open those nodes later.
Install it with npx or Docker. Everything lives in one local JSONL file, and MEMORY_FILE_PATH puts that file somewhere you control and back up. No accounts, no service costs.
The MCP steering group maintains it as a basic reference implementation. It works well for personal context like preferences and project facts. Need semantic search, sync across machines or memory shared by a team? That is a different class of tool.
Free, with memory in one local file. Set MEMORY_FILE_PATH so that file lives somewhere you back up.
**Free tier:** Free and open source (the project is moving from MIT to Apache 2.0). No underlying service. **Self-hosted:** Runs locally through `npx` or Docker. The graph is stored in a JSONL file, `memory.jsonl` in the server directory by default; set `MEMORY_FILE_PATH` to choose the location. **Paid:** None.
Sequential Thinking gives the model a structured scratchpad. Instead of answering in one shot, it records numbered thoughts, revises earlier ones, branches into alternatives and adjusts how many steps it expects to need.
One npx or Docker entry sets it up, and DISABLE_THOUGHT_LOGGING silences the thought log. The server is local and free, but every extra thought is another tool call your model has to generate, so answers cost more tokens.
The MCP steering group maintains it as a reference implementation. It earns its place on migrations, production debugging and architecture tradeoffs where course correction matters. Models with strong built-in reasoning get less out of it.
The server is free, but each thought step is another tool call, so hard problems cost more model tokens.
**Free tier:** Free and open source (the project is moving from MIT to Apache 2.0). No underlying service. **Self-hosted:** Runs locally through `npx` or Docker. Set `DISABLE_THOUGHT_LOGGING=true` to stop it logging thought content. **Paid:** None from the server. Each thought step is a tool call the model generates, which adds to your model usage.
Time fixes a small but real gap: models do not know what time it is. It exposes two tools, one for the current time in any IANA timezone and one for converting a clock time between zones, and detects your system timezone automatically.
Run it with uvx mcp-server-time or Docker. Override the detected zone with --local-timezone if you need to. Nothing to pay.
The MCP steering group maintains it as a reference implementation. It does one job: conversion takes an HH:MM time, not a full date, and there is no calendar logic. Cheap to add when deadlines or cross-timezone coordination come up in your prompts.
Free, local, no service. It converts clock times between zones, not full dates.
**Free tier:** Free and open source (the project is moving from MIT to Apache 2.0). No underlying service. **Self-hosted:** Runs locally with `uvx`, `pip` or Docker. System timezone is detected automatically and can be overridden with `--local-timezone`. **Paid:** None.
Context7 pulls current library documentation into your coding agent so it stops inventing API signatures from two-year-old training data. It covers a wide range of popular libraries, pre-chunked for LLM consumption, so the agent gets the relevant section instead of an entire documentation site. Setup changed this summer. Running the ctx7 setup command handles OAuth, generates an API key, and installs either the CLI-plus-Skills path or classic MCP mode, with flags to target Cursor, Claude, or opencode. Manual configuration against the hosted server with a bearer token still works. It runs without a key, but you share a rate limit with everyone else who skipped that step, and the key is free. Maintained by Upstash, MIT licensed, and the server costs nothing. Install it. The catch is coverage: Context7 can only serve what it has indexed, so a brand-new or obscure library leaves your agent guessing exactly like before.
Free, but run the ctx7 setup and take the free API key. Without one you share a rate limit with every anonymous user.
### Free Tier The MCP server and the ctx7 CLI are MIT licensed and free. The hosted documentation service is free to query, with a free API key available from the Context7 dashboard for higher rate limits. ### Self-Hosted Not the normal path. The value is Upstash's indexed and chunked documentation corpus, which is served from their infrastructure. You run only the local client or point your MCP client at the hosted endpoint. ### Paid No paid tier for the connector. Context7 lists an Enterprise option on its site for organizations that want their own private documentation indexed, priced on request.
GitHub's official MCP server gives your AI assistant direct access to repos, issues, PRs, code search, and file operations. No scraping, no workarounds. You talk to your assistant, it talks to GitHub. Setup is one config block plus a personal access token. Maintained by GitHub themselves, so it stays current with API changes. Covers the operations you actually use daily: creating branches, reviewing diffs, searching code across repos, managing issues. The scope is broad enough that it replaces most of your tab-switching. Worth installing immediately if you use GitHub. This is the one MCP server that pays for itself on day one. The catch: it can only do what the GitHub API allows, so anything requiring the web UI (like editing Actions workflows visually) is still manual.
Free and open source. Requires a GitHub account (free tier works).
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Connects your AI assistant to AWS VPC networking. You can inspect VPCs, subnets, security groups, route tables, and network ACLs. Useful for debugging connectivity issues or auditing your network configuration without clicking through the console. The MCP server is free and open source. VPC resources themselves have no hourly cost (NAT Gateways and some endpoints do). Setup just requires AWS credentials with the right IAM permissions, which most AWS developers already have configured. Maintained by AWS Labs. If you spend time troubleshooting security group rules or tracing packet paths through your VPC, having this in your AI assistant is genuinely useful. Pure read operations, low risk.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying AWS service may have its own pricing.
Connects your AI assistant to Finch, AWS's open source container development tool (an alternative to Docker Desktop). You can build images, manage containers, and run compose stacks. Think of it as Docker commands you can describe in natural language. The MCP server is free and open source. Finch itself is free and open source with no licensing concerns, unlike Docker Desktop's per-seat pricing for larger companies. Setup requires Finch installed on your machine. Maintained by AWS Labs. Good fit if you've already switched from Docker Desktop to Finch or are evaluating the move. The MCP integration makes container operations conversational. If you're happy with Docker Desktop, no reason to switch just for this.
## Free Tier The MCP server and Finch are both completely free and open source. ## Self-Hosted Runs locally on your machine. ## Paid No paid tier. No licensing fees.
Prometheus MCP lets your AI assistant query AWS Managed Prometheus metrics using PromQL. Ask questions about your infrastructure and get real data back. Setup needs your Prometheus endpoint and AWS credentials. AWS Managed Prometheus bills on ingested samples and queries. The MCP is free. Maintained by AWS Labs. Having your AI assistant understand your metrics is powerful for debugging and capacity planning. The catch: you need to know PromQL or trust your AI to write correct queries. Bad queries against production metrics can be slow and expensive.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
This server gives your AI assistant direct access to AWS Identity and Access Management. Query policies, roles, and users without switching to the console. Setup is just AWS credentials in your MCP config. IAM itself is free. No cost for the service or the MCP. Maintained by AWS Labs, so it tracks API changes closely. Useful for anyone managing AWS permissions. Debugging "why can't this role do X" becomes a conversation instead of a console scavenger hunt. The catch: you are giving your AI assistant read access to your IAM config. Make sure your credentials are scoped appropriately.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Connects your AI assistant to AWS HealthOmics for genomics and bioinformatics workflows. The model can manage reference stores, sequence stores, annotation stores, and run genomics workflows. Highly specialized for life sciences teams. You get workflow management, data store operations, and run monitoring across HealthOmics services. Requires AWS credentials with HealthOmics permissions. The MCP server is free, HealthOmics pricing is usage-based. Maintained by AWS Labs. This is extremely niche. If you are doing genomics work on AWS, it is a welcome addition. For everyone else, it is irrelevant. Know your audience before installing.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Connects your AI assistant to Aurora DSQL, Amazon's distributed SQL database. The model can run queries, inspect schemas, and explore data across a globally distributed relational database. You get SQL query execution and schema inspection against Aurora DSQL clusters. Requires AWS credentials and a running DSQL cluster. The MCP server is free, Aurora DSQL pricing is based on compute and I/O usage. Maintained by AWS Labs. Aurora DSQL is still relatively new, so this server is early-stage but functional. If you are building on DSQL and want AI-assisted database work, it is currently the only option. Worth installing for DSQL users.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
This integration connects your AI assistant to Amazon Bedrock Knowledge Bases. Query your ingested documents using RAG (retrieval-augmented generation) directly from your editor. Setup requires a Bedrock Knowledge Base already configured with data sources. Bedrock Knowledge Base pricing includes storage, indexing, and retrieval charges. The MCP is free. AWS Labs maintains it as a reference integration. Powerful if you have company docs indexed in Bedrock and want your AI assistant to search them. The catch: this is a reference implementation, not a production-ready server. Expect to customize it for your use case.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Connects your AI assistant to AWS HealthImaging, the service for storing, accessing, and analyzing medical images (DICOM format). You can search image sets, retrieve metadata, and browse imaging datastores from your editor. The MCP server is free and open source. HealthImaging charges per GB stored and per retrieval request. Setup requires an active HealthImaging datastore and IAM permissions, plus HIPAA compliance considerations for any production medical data. Maintained by AWS Labs. Extremely niche. This is for healthcare developers building on AWS HealthImaging who want AI-assisted exploration of their medical imaging pipeline. If you're in that space, it's a clean integration. Everyone else, skip it.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying AWS service may have its own pricing.
Connects your AI assistant to Amazon ECS for container management. You can check service health, inspect task definitions, view running tasks, and troubleshoot deployment issues. Covers both Fargate and EC2 launch types. The MCP server is free and open source. ECS itself is free (you pay for underlying Fargate or EC2 compute). Setup requires AWS credentials with ECS permissions. Be cautious with write operations, stopping tasks or updating services in production needs careful IAM scoping. Maintained by AWS Labs. If you manage ECS services and spend time checking deployments or debugging task failures, this puts that workflow in your editor. The ability to inspect stopped task reasons conversationally is worth the setup alone.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying AWS service may have its own pricing.
HealthLake MCP connects your AI assistant to AWS HealthLake, Amazon's FHIR-compliant health data store. You get read and write access to FHIR resources (patients, observations, conditions) directly from your editor. Setup requires AWS credentials and a running HealthLake data store, which is not trivial. HealthLake itself bills per request and per GB stored. Not cheap for experimentation. The MCP just wraps the API, so you still need the underlying service provisioned and configured. Worth it if you are already deep in HealthLake and want faster iteration on FHIR queries. Skip it if you are not in healthcare. The catch: HealthLake is one of AWS's pricier services, so testing against real data adds up fast.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Connects your AI assistant to Amazon MQ, which runs managed ActiveMQ and RabbitMQ brokers. Manage brokers, inspect queues, publish messages, and troubleshoot message flow without switching to the AWS console. Requires AWS credentials and an existing Amazon MQ broker. The MCP server is free. Amazon MQ pricing is per-hour based on broker instance size, starting around $0.03/hour for small instances. Maintained by AWS Labs. Solid pick if you manage message brokers on AWS and want AI-assisted operations. Not worth setting up Amazon MQ just to use this server.
Free MCP server. Amazon MQ charges per broker instance hour.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Connects your AI assistant to Amazon Q Business in anonymous mode. The model can query your Q Business application without user authentication, making it useful for internal knowledge retrieval and document search across your organization's indexed content. You get Q Business chat and query capabilities in anonymous access mode. Requires AWS credentials and an existing Q Business application. The MCP server is free, Q Business pricing is usage-based starting at $3/user/month. Maintained by AWS Labs. Useful if your org already runs Q Business and you want to pipe its knowledge into your AI workflow. Niche but practical for existing AWS enterprise setups.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Loads documents into your AI assistant's context. Supports PDF, HTML, Markdown, CSV, Excel, and plain text. You can point it at local files or URLs and get the content extracted and ready for analysis or summarization. The MCP server is free and open source with no underlying service costs. It runs entirely locally using Python-based document parsing. Setup is installing the server and its dependencies. No AWS account needed. Maintained by AWS Labs, but this one is AWS-agnostic. Handy utility for anyone who regularly feeds documents into AI conversations. Simple, does one thing well. Worth installing if you work with varied document formats.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally using Python-based document parsing. ## Paid No paid tier. No external service dependencies.
What it connects: AWS Systems Manager for SAP, exposing the full SAP lifecycle (register apps, run backups, push configs, manage HA events) through an MCP interface for your AI agent. What you can do with it: Backup SAP databases, configure runtime settings, react to high-availability events, and manage registered SAP applications via conversational commands. Setup needs AWS credentials, SSM for SAP registered apps, and a Python environment. What the service costs: The MCP is free. AWS Systems Manager for SAP has per-managed-instance pricing on top of your SAP workload costs. Who maintains it: AWS Labs, official. Apache 2.0, part of the awslabs/mcp monorepo. Verdict: Worth installing if you run SAP on AWS. Otherwise, nothing here matters.
**Free tier:** Fully Apache 2.0, no paid tier. **Self-hosted:** Runs as a local Python MCP server. Requires AWS credentials and SSM for SAP registered applications. **Paid:** N/A. The underlying AWS Systems Manager for SAP has its own per-managed-instance pricing.
Well-Architected Security MCP gives your AI assistant access to AWS's security review framework. Run security posture checks, review configurations against best practices, and surface risks. Setup is AWS credentials with appropriate permissions. The underlying Well-Architected Tool is free. The MCP is free. AWS Labs maintains it. Handy for getting a quick security gut-check on your AWS setup without opening the console. The catch: it reviews against AWS best practices, not your specific threat model. Passing does not mean secure. It means you follow the checklist.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Redshift MCP connects your AI assistant to Amazon Redshift, AWS's data warehouse. Query tables, explore schemas, and analyze data directly from your editor. Config is your Redshift connection details plus AWS credentials. Redshift pricing varies by cluster type. Serverless bills per query. The MCP is free. AWS Labs maintains it. Strong pick for data teams already on Redshift who want AI-assisted analysis without context-switching. The catch: Redshift queries can get expensive fast. No built-in guardrails on query cost, so watch your serverless RPU usage.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Connects your AI assistant to AWS AppSync GraphQL APIs. The model can manage schemas, resolvers, data sources, and run GraphQL operations. Useful for building and debugging AppSync backends without constantly switching between console tabs. You get full AppSync management: schema updates, resolver configuration, data source setup, and query execution. Requires AWS credentials. The MCP server is free, AppSync charges per query and per real-time connection. Maintained by AWS Labs. If AppSync is your GraphQL layer, this collapses a lot of console clicking into AI-assisted commands. Solid productivity boost for teams deep in the AppSync ecosystem.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Connects your AI assistant to Amazon ElastiCache for Redis and Memcached. You can check cluster health, inspect node status, review replication groups, and monitor cache performance metrics. Useful for diagnosing cache hit rates and connection issues. The MCP server is free and open source. ElastiCache pricing starts around $0.017/hour for the smallest node. The server needs network access to your ElastiCache cluster, which lives inside a VPC, so connectivity setup is the main hurdle. Maintained by AWS Labs. Caching problems are notoriously hard to debug. If you're running ElastiCache and regularly troubleshoot eviction rates or replication lag, having this in your editor saves console tab-switching. VPC access requirement adds setup friction.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying AWS service may have its own pricing.
Connects your AI assistant to CloudWatch metrics, logs, and alarms. You can query log groups, pull metric data, check alarm states, and search through log streams. This is the workhorse of AWS observability, now accessible conversationally. The MCP server is free and open source. CloudWatch charges per metric, per log GB ingested, and per query. Most AWS accounts already generate CloudWatch data, so there's no new cost from the MCP server itself. Setup is just AWS credentials. Maintained by AWS Labs. One of the most useful MCP servers in the entire collection. Every AWS developer checks CloudWatch logs. Querying them by describing what you're looking for instead of writing CloudWatch Insights syntax is a real time saver.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying AWS service may have its own pricing.
SageMaker MCP connects your AI assistant to Amazon SageMaker's ML platform. Manage training jobs, endpoints, and model artifacts from your editor. Setup requires AWS credentials and SageMaker resources already provisioned. SageMaker pricing is complex: compute by the hour, storage by the GB, endpoints by uptime. The MCP itself is free. AWS Labs maintains it. Useful for ML engineers who live in SageMaker and want to check job status or deploy models without the console. The catch: SageMaker's API surface is enormous. The MCP exposes a subset, so you will still hit the console for advanced workflows.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Connects your AI assistant to Amazon Keyspaces, AWS's managed Cassandra-compatible database. Create tables, run CQL queries, and manage schemas through natural language instead of wrestling with the Cassandra CLI. Setup needs AWS credentials and a Keyspaces endpoint. If you already use Keyspaces, this slots right in. The MCP server is free. Keyspaces pricing is pay-per-use based on reads, writes, and storage, with a modest free tier. Maintained by AWS Labs. Useful if you're already running Keyspaces workloads. For everyone else, there are simpler database MCP servers to start with.
Free MCP server. Amazon Keyspaces uses pay-per-use pricing with a free tier.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Postgres MCP connects your AI assistant to Aurora PostgreSQL databases. Run queries, inspect schemas, list tables. Setup is a connection string in your MCP config. Aurora Postgres pricing is usage-based. The MCP is free. AWS Labs maintains it with read-only query defaults. Essential if you are doing any development against Aurora Postgres. Having your AI understand your schema changes the conversation entirely. The catch: optimized for Aurora Postgres, not vanilla PostgreSQL. It works with standard Postgres, but Aurora-specific tooling is the focus.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Connects an AI assistant to an Oracle Database running on AWS RDS. It defaults to read-only transactions, allows writes only behind a flag, and supports schema introspection. It lets an assistant query and explore an Oracle schema without hand-written SQL. A nice touch: it uses python-oracledb thin mode, so there is no painful Oracle Instant Client to install. Setup is still heavy, needing AWS credentials, RDS details, and logins in AWS Secrets Manager. The server is free and Apache-2.0, running against an Oracle RDS instance you already pay for. Maintained by AWS in the official awslabs/mcp repository. Worth it for AWS shops running Oracle on RDS. It is scoped to RDS only, so a self-managed Oracle box is out of reach.
**The MCP server:** Free and Apache-2.0. **Underlying service:** Oracle Database on AWS RDS, which you already pay for (RDS plus Oracle licensing, license-included or BYOL). The server adds no incremental cost beyond your existing RDS instance.
Connects your AI assistant to CloudWatch Application Signals, AWS's application performance monitoring layer. You can check service health, review SLO compliance, and investigate performance anomalies across your instrumented services. The MCP server is free and open source. Application Signals pricing is based on the number of monitored operations. Setup requires your applications to be instrumented with the CloudWatch agent or OpenTelemetry, which is the real lift here. Maintained by AWS Labs. Only useful if you've already adopted Application Signals for APM. If you're using Datadog, New Relic, or even plain CloudWatch metrics, this won't apply. For teams already invested in the AWS observability stack, it's a natural extension.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying AWS service may have its own pricing.
Connects your AI assistant to Amazon Neptune graph databases. The model can inspect schema, check cluster status, and run openCypher or Gremlin queries against your graph data. Graph databases are notoriously hard to query by hand, so letting AI help is a genuine productivity win. You get schema inspection, status monitoring, and query execution. Requires AWS credentials and a running Neptune cluster. The MCP server is free, but Neptune pricing starts around $0.10/hour for the smallest instance. Maintained by AWS Labs, the official open source arm of AWS. If you run Neptune and want AI-assisted graph exploration, this is the only game in town. Solid choice for existing Neptune users.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Connects your AI assistant to AWS Glue and Athena for data processing. The model can build ETL jobs, run SQL queries against S3 data, manage crawlers, and monitor pipeline execution. Turns your AI into a data engineering co-pilot. You get Glue job management, Athena query execution, crawler configuration, and pipeline monitoring. Requires AWS credentials. The MCP server is free, Glue and Athena charge based on compute time and data scanned. Maintained by AWS Labs. If your data stack runs on Glue and Athena, this is a significant productivity multiplier. Writing Glue jobs and Athena queries with AI assistance beats the console experience by a wide margin.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying service may have its own pricing.
Connects your AI assistant to the AWS Pricing API so you can look up on-demand pricing for any AWS service. Ask about EC2 instance costs, compare RDS pricing across regions, or estimate what a Lambda workload will run you. The MCP server is free and open source. The Pricing API itself is free to call. Setup is minimal, just AWS credentials. This is one of the simplest MCP servers in the AWS Labs collection. Maintained by AWS Labs. Genuinely useful for anyone building on AWS who regularly checks pricing. Beats opening the pricing calculator in a browser tab. Quick to install, no ongoing maintenance.
## Free Tier The MCP server itself is completely free and open source. ## Self-Hosted Runs locally or on your own infrastructure. ## Paid No paid tier. The underlying AWS service may have its own pricing.
Connects your Chrome browser to AI assistants via MCP. Claude can take screenshots, fill forms, manage bookmarks, and search across your open tabs using a local vector database. Twenty-plus tools, all running locally. No data leaves your machine. Install the Chrome extension and the bridge package, configure your MCP client, and your AI can see what you see. Setup takes about 10 minutes. Maintained by hangwin (community). Free, MIT licensed, no API keys or subscriptions. If you want your AI assistant to actually interact with your browser, this is the one to install.
## Free Tier The MCP server itself is completely free and open source. Install it, configure your MCP client, and use it without any cost. ## Self-Hosted Runs locally or on your own infrastructure. No external dependencies beyond the service it connects to. ## Paid No paid tier. The MCP server is free. The underlying service may have its own pricing.