AWS MCP Servers
★ 9.4k · GitHub →AWS Labs' suite of MCP servers — documentation, IaC, cost, databases, and service APIs for agents.
What is AWS MCP Servers?
awslabs/mcp is not one server but a monorepo of dozens of specialized MCP servers, each wrapping a slice of AWS: documentation and knowledge retrieval, infrastructure-as-code (CDK, CloudFormation, SAM), containers (EKS, ECS), databases (DynamoDB, Aurora, DocumentDB, Redshift and more), AI/ML services (Bedrock Knowledge Bases, SageMaker), and cost analysis. Most install with a single uvx command and authenticate through whatever AWS credentials you already have — profiles, IAM roles, environment variables.
Speaking as someone who works on AWS daily: the documentation/knowledge server and the cost server earn their keep immediately. AWS docs are vast and the model's memory of service quirks is unreliable, so letting the agent query current documentation kills a whole class of "that parameter was renamed two years ago" errors. The CDK and IaC servers are genuinely useful for scaffolding constructs with current best practices baked in.
The honest caveats: quality varies noticeably across servers, because they are maintained by different AWS teams at different levels of care — some are polished, some feel like proofs of concept. The collection also moves fast: SSE transport was dropped in favor of Streamable HTTP across the suite, servers get added and consolidated, and AWS now points production workloads toward its newer agent tooling. Treat each server as its own product and read its README before trusting it.
Key capabilities
- ▸ Dozens of specialized servers spanning docs, IaC (CDK/CloudFormation/SAM), containers, databases, AI/ML, and cost
- ▸ One-command install via uvx for most servers; Docker images and one-click installers for popular clients
- ▸ Authenticates with your existing AWS credentials — profiles, IAM roles, AWS_REGION — no separate account
- ▸ Documentation and knowledge servers ground the agent in current AWS docs instead of stale training data
- ▸ Cost-analysis tooling lets the agent reason about pricing and spend before provisioning anything
When to use it
If your daily work touches AWS, start with two or three servers — documentation, cost, plus whichever matches your stack (CDK, EKS, a database) — and add more only when a task demands it. Loading many servers at once bloats the tool list and degrades the agent's tool selection. Skip the suite entirely if you are not on AWS, obviously, but also if you only need occasional AWS facts: web search plus official docs covers light usage without the credential surface. And be deliberate about which IAM permissions the agent-facing profile carries — several servers can mutate real infrastructure.
Alternatives
Frequently asked questions
How do I add the AWS documentation MCP server to Claude Code?
Run `claude mcp add aws-docs -- uvx awslabs.aws-documentation-mcp-server@latest` (uv must be installed). It needs no AWS credentials since it only reads public documentation, which makes it the safest first server from the suite to try. Servers that call AWS APIs — cost, CDK, databases — additionally read your AWS_PROFILE and AWS_REGION environment variables.
Do the AWS MCP servers need my AWS credentials?
Depends on the server. Documentation and knowledge servers work without credentials. Anything that touches your account — cost analysis, database servers, IaC deployment helpers — uses the standard AWS credential chain: named profiles, environment variables, or an assumed IAM role. Best practice is a dedicated profile with least-privilege permissions rather than your admin credentials.
Which AWS MCP server should I install first?
The documentation server. It requires no credentials, carries no write risk, and fixes the most common failure — the agent confidently citing outdated AWS API details. From there, add the cost server if you provision infrastructure, and the CDK or CloudFormation server if you write IaC. Resist installing the whole suite; a fat tool list makes the agent choose tools worse.
Star count updated 2026-07-02 · refreshed weekly from the GitHub API · browse all tools →