Skip to content
Nubent

/AI-native: AWS + Claude

MCP servers that connect your agents to internal systems and AWS

We design and build Model Context Protocol (MCP) servers so Claude and other agents can query and act on your APIs, databases, tickets and observability. Every server ships with authentication, least-privilege permissions and audit logging, deployed in your AWS account on Lambda, ECS or Amazon Bedrock AgentCore.

/Diagnosis

Signs it is time

Your agents can't see your systems

Claude gives good general answers, but it can't query your CRM, your tickets or your database, so the team copies and pastes context by hand.

The same integration, built per tool

Every assistant or agent has its own integration with the same APIs, with duplicated code and different permissions in each one.

Unvetted third-party MCP servers

Engineers install public MCP servers with personal tokens, and nobody knows what data they expose or what actions they allow.

Security wants traceability

Before letting an agent touch internal systems, your security team wants identity, scoped permissions and a log of every call.

/What we deliver

Concrete deliverables, in your repository

  • Tool and resource design

    A catalog of MCP tools and resources per system, with names, descriptions and schemas designed so the model uses them well.

  • Production MCP servers

    TypeScript or Python code built on the official MCP SDKs, with tests, input validation and error messages the agent can act on.

  • Authentication and permissions

    OAuth or IAM depending on the client, least-privilege roles per tool and a clear split between read and write operations.

  • Audit logging and observability

    Structured logs of every call with identity, tool and parameters, CloudWatch metrics and alarms on errors or abuse.

  • AWS deployment with Terraform

    Infrastructure on Lambda, ECS Fargate or AgentCore, with private networking, secrets in Secrets Manager and a CI/CD pipeline.

/How we work

Our process

  1. Map systems and use cases

    We identify which questions and actions the agents need, which systems answer them and what risk each one carries.

  2. Design the MCP interface

    We define tools and resources, schemas, permissions and the authentication model, and validate them with Claude on real cases.

  3. Build and test

    We implement the servers in your repository by pull request, with automated tests and evaluation of how the agent uses them.

  4. Deploy and operate

    We deploy in your AWS account, connect the approved MCP clients and monitor usage, errors and cost.

What an MCP server is and what it is for

Model Context Protocol (MCP) is an open protocol, created by Anthropic and now hosted by the Linux Foundation's Agentic AI Foundation, that standardizes how an AI agent connects to external tools and data. An MCP server exposes tools, resources and prompts; an MCP client, such as Claude Code or your own agent, discovers and uses them.

The practical advantage is reuse. Instead of writing a separate integration for each assistant, you build one MCP server per system and connect it to every client that needs it. Your ticketing system, your customer database or your observability platform becomes available to any authorized agent, under the same rules.

Designing tools a model actually uses well

Mirroring your API endpoint by endpoint rarely makes a good MCP server. Expose a hundred cryptically named endpoints and the model picks the wrong ones, burns context and makes mistakes. We design tools around the agent's tasks: few of them, with clear names, precise descriptions, validated parameters and compact responses that fit well in context.

We also take care with errors. A message like "customer not found, try searching by email" lets the agent correct its next step, while a generic error leaves it guessing. We test each server with Claude on real cases and refine descriptions and schemas based on how the model uses them.

MCP security: authentication, permissions and auditing

An MCP server gives an agent the ability to read data and take actions, so we treat it like any sensitive API. We authenticate every client with OAuth or IAM, assign least-privilege permissions per tool and separate read operations from write operations, which can require human confirmation.

Every call is logged with the identity of the user or agent, the tool invoked and its parameters. Those logs go to CloudWatch and can be correlated with CloudTrail. We also review the third-party MCP servers your team wants to use, to decide which ones are approved and with what scope.

  • Read-only by default; write access only where needed, and with limits.
  • Secrets in AWS Secrets Manager, never in code or local configuration.
  • Input validation to reduce the risk of prompt injection reaching your systems.

Deploying MCP servers on AWS

For lightweight servers with spiky traffic, AWS Lambda behind API Gateway is a simple, low-cost option. When a server needs persistent connections, more memory or access to private networks, we use Amazon ECS on Fargate. Everything is defined in Terraform, with networking, roles and secrets versioned in your repository.

Amazon Bedrock AgentCore adds building blocks designed for agents: AgentCore Runtime can host MCP servers and agents, and AgentCore Gateway provides a single entry point that groups tools and enforces authentication. We assess whether those managed capabilities make sense for you or whether Lambda and ECS are enough.

/FAQ

MCP servers: frequently asked questions

Another question? Write to hello@nubent.com and an engineer will answer.

What is the Model Context Protocol (MCP)?

MCP is an open protocol that defines how AI agents connect to external tools, data and services. Anthropic created it in 2024, and in December 2025 it moved to the Agentic AI Foundation, part of the Linux Foundation. Claude and many other tools support it today, which means the same server can be reused across different clients.

Should I build my own MCP server or use an existing one?

For well-known public services, there is often an official MCP server worth evaluating first. For your internal systems, or when you need control over permissions, auditing and response format, your own server is usually the better choice. We review what already exists, decide with you and build only what adds value.

Where do you deploy an MCP server on AWS?

The usual options are AWS Lambda with API Gateway for light workloads, Amazon ECS on Fargate for servers with long-lived connections or private network access, and Amazon Bedrock AgentCore, which provides a managed runtime and gateway for agents and MCP tools. We choose based on traffic, latency, network requirements and how clients authenticate.

Is it safe to connect an AI agent to internal systems with MCP?

It is, if you design it with the same controls as any sensitive API: authentication for every client, least-privilege permissions per tool, separation of read and write, input validation and a log of every call. For actions with real impact, we add human confirmation. The risk grows when unvetted servers get installed with broad personal credentials.

Do MCP servers work with Claude Code?

Yes. Claude Code is an MCP client and can connect to local or remote servers defined in the project configuration. That lets an engineer look up tickets, documentation or metrics from the same working session. The same servers also work for agents built with the Claude Agent SDK or other frameworks.

Next step

Let's talk about your MCP servers project

30 minutes with an engineer. Concrete next steps, whether or not you work with us.