Search approved knowledge
Give an assistant access to selected documents, policies or records, with links back to the source.
Custom MCP server development
Give AI controlled access to real systems
Astraeus designs and builds Model Context Protocol servers that give AI assistants a clear, controlled way to use the tools and information a business already relies on.
Scope an MCP server
An MCP server sits between an AI assistant and a business system. It can expose approved information as resources, provide reusable prompts or offer tightly scoped tools such as finding a CRM record, searching internal guidance or creating a draft task.
The useful work is not the protocol on its own. It is deciding what should be available, what must stay private, which actions need approval and how somebody can reconstruct what happened when a call fails or produces the wrong result.
The sequence
How the work moves
- 01
Define the boundary
Name the users, systems, data, actions and decisions that must remain human.
- 02
Design the interface
Specify the resources, tools, schemas, authentication and error behaviour before implementation.
- 03
Build and test
Implement the server, validate inputs and outputs, and run expected, failure and permission test cases.
- 04
Deploy and hand over
Document configuration, credentials, logs, support ownership and the process for changing a tool safely.
What a custom MCP server can do
Read business systems
Retrieve the permitted fields from HubSpot, Attio, Slack, databases or another documented API.
Prepare bounded actions
Create a draft CRM update, task, message or workflow request for a person to approve.
Join several tools
Present one stable interface when the assistant needs information from more than one underlying system.
What the build includes
Tool and resource design
Names, descriptions and schemas written so the AI client can understand when and how each capability should be used.
Authentication and permissions
Credentials stay outside the model context, access is limited and sensitive actions require the right identity or approval.
Validation and failure handling
Inputs are checked, errors are clear and unsafe or ambiguous requests stop without carrying out the action.
Testing and evaluation
Expected calls, denied calls, malformed inputs and model misunderstandings are tested before release.
Logging and operating records
Enough information is recorded to investigate what ran, what failed and which user or assistant requested it.
Deployment and handover
Configuration, runbooks and change notes let the client understand how the server is hosted and maintained.
Common MCP projects
These are implementation patterns, not fixed packages.
CRM assistant
Search approved HubSpot or Attio records, prepare updates and require a person to approve changes.
Internal knowledge server
Expose selected SharePoint, Drive, Notion or database content with source links and access checks.
Slack or Teams operations
Let an approved assistant find agreed channel information or prepare a message, task or escalation for review.
Public content server
Publish a read-only MCP surface for website pages, documentation or another controlled public source.
The honest boundary
A custom MCP server is not automatically better than an API, webhook or built-in connector. Astraeus recommends MCP when several AI clients need a stable, discoverable interface or when the controls around tool use justify the extra layer. Sensitive or irreversible actions still need explicit approval.
Direct answers
Common questions
What is MCP in plain English?
Model Context Protocol is a standard way for an AI assistant to discover and use approved information or actions. The server describes what is available and handles the actual connection to the business system.
Do you build custom MCP servers?
Yes. Astraeus can scope, build, test, deploy and document a custom server, including authentication, permissions, validation, logs and handover.
Can it connect Slack, HubSpot or Attio?
Potentially, yes. The exact design depends on the product API, the client account, the required records or actions and the permissions available.
Does it work with ChatGPT, Claude or other AI clients?
It can work with MCP-compatible clients. Support and authentication options change, so compatibility is checked against the intended client before the build is agreed.
Where is the server hosted?
That depends on the data, security and operating requirements. It may run in client-controlled infrastructure or an agreed managed environment, with credentials stored outside the model context.
When should we use a normal API instead?
Use the simpler option when one application needs one predictable connection. MCP becomes useful when AI clients need discovery, several related tools or a consistent interface across changing systems.
Start with the system and the boundary.
Tell us which AI client needs access to which data or action. We will say whether MCP is justified and what a safe first version would include.
Scope an MCP server