H
    Hestur
    MCPModel Context Protocol — Custom Server Development

    Custom MCP Servers — Your Systems as AI Tools.

    MCP is the standard that lets Claude, ChatGPT, Cursor, and Lovable call your internal systems as tools. We build production MCP servers with RBAC, OAuth, audit logging, and real-world system integrations — deployed in 4–8 weeks.

    What MCP is

    The USB-C for AI tools.

    Before MCP, every AI integration was a custom API connector — Slack had a different format than Salesforce, which was different from your internal database. Every connection was unique engineering work.

    MCP (Model Context Protocol, released by Anthropic) standardises how AI agents call external tools. An AI agent that speaks MCP can use any MCP server — regardless of what system it connects to. Build once, use with any compliant client.

    Claude, ChatGPT, Cursor, Lovable, and most production AI frameworks now support MCP. Which means a CRM MCP server you build today works across all of them — not just Claude.

    Architecture

    AI Agent (Claude, ChatGPT, Cursor, Lovable)
            │
            │  MCP protocol (JSON-RPC over stdio / HTTP/SSE)
            ▼
      ┌─────────────────────────────────┐
      │        MCP Server               │
      │  ┌─────────────────────────┐   │
      │  │  Tool Registry          │   │
      │  │  search_contacts()      │   │
      │  │  create_lead()          │   │
      │  │  update_deal()          │   │
      │  └─────────────────────────┘   │
      │  ┌─────────────────────────┐   │
      │  │  Auth Layer (OAuth/JWT) │   │
      │  │  RBAC enforcement       │   │
      │  │  Audit logging          │   │
      │  └─────────────────────────┘   │
      └──────────────┬──────────────────┘
                     │
            ┌────────┴────────┐
            │                 │
       Salesforce CRM    Internal DB
       (your existing    (your existing
        systems)          systems)

    MCP is supported by

    Claude (Desktop + API)ChatGPT (Plugins/MCP)CursorLovableWindsurfZed EditorCustom LangChain/LangGraph agentsCrewAIn8n (MCP node)

    What we build

    Six common MCP server patterns.

    Each MCP server exposes a defined set of tools — functions the AI agent can call. The server handles auth, rate limiting, and logging. Your systems stay exactly where they are.

    CRM MCP Server

    2–3 weeks

    Salesforce, HubSpot, Pipedrive, Zoho

    search_contacts, create_lead, update_deal, log_activity, get_pipeline_stage

    Claude or ChatGPT looks up a contact mid-conversation, creates a deal, updates a stage — without the AI having direct API access to your CRM.

    Internal Docs MCP Server

    2–4 weeks

    Confluence, Notion, SharePoint, Google Drive

    search_docs, get_page, list_spaces, search_by_tag, get_recent_changes

    Cursor or Claude Desktop can search your internal knowledge base during coding or writing sessions — without manual copy-paste of context.

    Jira / Project Management MCP Server

    1–3 weeks

    Jira, Linear, GitHub Issues, Asana

    create_issue, search_issues, update_status, assign_ticket, get_sprint_board

    AI agents create tickets from conversation, update sprint boards, and pull issue context into code reviews — all through natural language.

    Salesforce Enterprise MCP Server

    4–6 weeks

    Salesforce (Sales Cloud, Service Cloud, CPQ)

    query_soql, get_opportunity, update_contact, run_report, get_account_hierarchy

    Sales agents get full Salesforce context in Claude — account history, open opportunities, renewal dates — without copying data into the chat.

    Custom Database MCP Server

    3–6 weeks

    PostgreSQL, MySQL, MongoDB, BigQuery, Snowflake

    query_safe (read-only SQL), search_records, aggregate_data, export_results

    AI analysts query your data warehouse in natural language through a safe, read-only MCP interface — no direct DB access, no SQL injection risk.

    E-commerce Operations MCP Server

    2–4 weeks

    Shopify, WooCommerce, BigCommerce + fulfilment APIs

    get_order, update_fulfilment_status, check_inventory, create_refund, search_products

    Customer service AI agents handle order lookups, refund creation, and inventory checks without needing a separate UI or human handoff.

    Security

    Production MCP security: RBAC, OAuth, and audit logging.

    Every MCP server we build includes the security layer your enterprise requires. These are not optional add-ons — they are in the default implementation.

    RBAC (Role-Based Access Control)

    Tool exposure is scoped per client and per role. A sales agent MCP server sees contacts and deals — not billing data. An analyst MCP server is read-only. Access boundaries are enforced at the server, not the client.

    OAuth 2.0 / PKCE Integration

    Users authenticate through your existing identity provider (Okta, Auth0, Azure AD, Google Workspace). The MCP server verifies JWTs before executing any tool call. No shared API keys, no static credentials.

    Tool-Level Audit Logging

    Every tool call is logged: which AI agent called it, which user triggered the call, what arguments were passed, what was returned, and the timestamp. Logs are immutable and exportable for compliance review.

    Rate Limiting and Quota Enforcement

    Tool call rates are capped per API key and per user. Burst protection prevents a runaway AI agent from hammering downstream APIs. Quotas are configurable per integration.

    No Direct System Access from LLMs

    The LLM calls the MCP server via a defined protocol — it never gets raw API keys, database credentials, or direct network access to your systems. The MCP server is the only authorised client of your internal APIs.

    Data Redaction Before Model Exposure

    Configurable PII and sensitive field redaction on MCP server responses. SSNs, card numbers, internal pricing, and other sensitive fields can be masked before data reaches the model context window.

    Delivery

    4–8 weeks from scoping to production.

    01Week 1

    Discovery & Scoping

    We map the tools the AI agent needs, the systems they connect to, authentication flows, and access control requirements. We define the tool schema and test against your API docs.

    02Week 2–3

    Server Build & Auth Integration

    MCP server implementation in TypeScript or Python. OAuth/JWT integration with your identity provider. Tool implementations with error handling and rate limiting.

    03Week 3–4

    RBAC, Logging & Security Review

    Role definitions applied to tool exposure. Audit log pipeline configured. Security review: injection testing, auth bypass testing, data redaction validation.

    04Week 4–8

    Integration Testing & Deployment

    End-to-end testing with your AI agent (Claude Desktop, API, or custom agent). Load testing at expected call volumes. Deployment to your infrastructure (hosted or on-prem).

    Deployment

    Hosted on your cloud, or fully self-hosted.

    AspectHosted (cloud account)Self-hosted
    Infrastructure ownershipWe host on your cloud account (AWS, GCP, Azure)You deploy on your own servers
    MaintenanceWe manage updates, uptime, and scalingYour team manages infrastructure
    Data residencyConfigurable — your cloud regionFull control — on-prem or air-gapped
    Cost$200–600/month cloud infra (billed to you)Infrastructure cost only (~$50–200/month)
    Setup timeIncluded in projectAdd 1–2 weeks for DevOps
    Best forTeams without dedicated DevOpsRegulated industries, on-prem requirement

    Ready to give your AI agents real-world access?

    30-minute scoping call. We map the tools your AI agent needs, identify the systems, and design the server architecture. Most MCP servers are scoped to a firm price in the first session.