AI Sales Agent + CRM
Evidence-first architecture for deal context, correspondence, AI output validation and controlled CRM actions.
Explore the reference →MCP · AI · CRM · API
Custom MCP Server Development for AI Agents
Connect ChatGPT, Claude and AI agents to your CRM, databases, APIs and internal services. I build secure MCP integrations around the workflows your business actually needs.
- Custom development
- Works with existing systems
- Controlled data access
- Support and maintenance
What MCP does for your business
Model Context Protocol (MCP) is an open protocol that connects AI applications to external tools and data sources. It provides a shared connection interface; your integration still defines business rules and permissions.
What is MCP server development?
MCP server development is the creation of a software interface that gives compatible AI applications controlled access to the tools, data and functions of external systems.
Without an MCP integration
An AI assistant can answer questions, but cannot access your private CRM without a separate connection. People have to find and copy the context themselves.
With an MCP integration
A compatible AI client can call approved tools to find customers, review deals, retrieve orders and prepare reports. Write access is enabled only for agreed workflows.
Access is restricted by user permissions, security policies and server configuration. MCP does not bypass authorisation or give an AI agent unrestricted control of your business.
Connect the systems you already use
Integrations use available official APIs and agreed interfaces. Discovery checks permissions, subscriptions, rate limits and support for each required operation.
CRM
amoCRM, Bitrix24 and custom platforms: customers, deals, tasks and interaction history.
ERP
Orders, inventory, finance and production workflows within the scope of your system’s API and permissions.
Databases
PostgreSQL, MySQL and other databases through restricted views and parameterised queries.
Business APIs
REST, GraphQL and internal services with input validation and controlled network access.
Documents and RAG
Search company documentation and knowledge bases while respecting source permissions.
Messaging
Telegram, WhatsApp and other platforms, limited to operations available through their official APIs.
Find messages, summarise conversations and draft replies through authorised mail interfaces.
E-commerce
Products, orders, stock and analytics with permissions scoped by store, role and action.
From answering questions to getting work done
Start with an outcome that helps an employee, then define narrowly scoped tools. Reading data and changing it have separate permissions and controls.
READ ACCESS
Deal analysis
An agent reviews deal stages, tasks and correspondence to identify stalled opportunities.
READ ACCESS
Reports in plain language
A manager requests a summary for agreed metrics and a specific reporting period.
READ ACCESS
Documentation search
An employee finds the relevant policy and the sources supporting the answer.
READ ACCESS
Customer briefings
A sales representative receives a summary of orders, conversations and commitments.
WRITE / APPROVAL REQUIRED
CRM task creation
The agent prepares a task. Creation happens after human approval and a server-side permission check.
READ ACCESS
Orders and inventory
An agent analyses product availability, order trends and potential discrepancies.
WRITE / APPROVAL REQUIRED
Internal service workflows
An agent calls agreed service operations; actions with side effects go through explicit controls.
WRITE / APPROVAL REQUIRED
Proposed data changes
A person reviews and approves proposed changes, then the server checks the current record before writing.
How the MCP integration fits together
The MCP client runs inside the AI application. The server exposes tools and enforces access checks before reaching business systems. This diagram shows logical boundaries, not a requirement to deploy every step as a separate service.
- AI application / AI agentA compatible client: ChatGPT, Claude or your own application
- MCP ClientConnection component inside the AI application
- MCP ServerTools · Resources · Prompts
- Access control and authorisationPermissions · input validation · audit · rate limits
- CRM / ERP / API / Database / RAGOnly approved data and operations
MCP Tools
Focused operations such as finding a customer, fetching an order or preparing a task, with clear input and output schemas.
MCP Resources
Addressable data and documents with access policies. Availability depends on the AI client.
MCP Prompts
Reusable prompt templates where supported by the client. They do not replace server-side checks.
Authentication and operations
HTTPS, scoped tokens, audit trails, timeouts and rate limits. Streamable HTTP for remote access; stdio where a local process is appropriate.
Security for MCP integrations
MCP does not automatically make an integration secure. Protection comes from the architecture, correct authorisation, business rules and server-side validation of every call.
Authentication and roles
OAuth 2.1 for applicable HTTP integrations, token validation and object-level permissions. Other environments use a mechanism selected for the threat model.
Least privilege
Read-only access wherever writing is unnecessary; separate roles, tools and service accounts for different workflows.
Isolation and secret management
A restricted integration layer protects internal services. Secrets stay on the server. Arbitrary URLs, SQL and shell commands are not passed through for execution.
Audit and request limits
Record the actor, operation and outcome with sensitive values masked. Bound call frequency, runtime and data volume.
Human approval for actions
Sensitive actions require approval. The server rechecks permissions, parameters and record state before writing and accounts for retries.
Prompt injection defences
Treat email and document content as untrusted data. It cannot grant a tool extra privileges. Validate arguments and outputs, and test adversarial workflows.
Relevant engineering experience
These references demonstrate AI, CRM and RAG architecture. They are not presented as completed commercial MCP deployments.
AI Beauty Booking
AI-native scheduling and CRM architecture with knowledge retrieval and workflow automation.
Explore the reference →RAG Content Pipeline
Public architecture and code examples: LLM classification, Qdrant, quality checks and controlled publishing.
Explore the reference →IdeaRadar
A public product for discovering and evaluating software projects, showing data processing and product engineering.
Explore the reference →How I develop your MCP server
Review systems and APIs
Check documentation, supported operations, constraints and data sources.
Define the agent’s job
Agree on users, workflows, useful outcomes and acceptance criteria.
Design tools and resources
Define contracts, parameter schemas, read/write boundaries and access policies.
Build the server
Implement adapters, explicit errors, timeouts and safe retries.
Configure access controls
Set up roles, tokens, approvals, isolation and request limits.
Test the integration
Verify chosen AI clients, API failures, permissions and attempts to bypass restrictions.
Deploy
Prepare hosting, HTTPS, secrets, monitoring and an update procedure.
Document and maintain
Hand over client setup, tool contracts and the support process.
What does custom MCP server development cost?
Scope depends on the number of systems and tools, API quality, access controls, hosting and testing requirements.
Single-system MCP integration
One API or data source, a focused set of tools and a clearly defined workflow.
Pricing is provided after reviewing your requirements.
Get a project estimate →Enterprise MCP server
Multiple systems, authentication, role-based access, audit trails and operational limits.
Pricing is provided after reviewing your requirements.
Get a project estimate →AI agent + MCP
An AI interface, tool integrations and end-to-end workflow automation with action controls.
Pricing is provided after reviewing your requirements.
Get a project estimate →A stack that fits your infrastructure
I select the stack after discovery. The MCP server can run as a Node.js or Python service while business logic stays in PHP/Laravel. SDK and protocol-version support are checked for your project. Audits, optimisation and ongoing maintenance are also available for existing MCP servers.
- PHP / Laravel
- Node.js
- Python
- PostgreSQL / MySQL
- Redis
- REST / GraphQL
- Docker
- OAuth
- MCP
- RAG
- LLM API
Technical references: MCP architecture · Authorisation · Security practices · ChatGPT connections
MCP development: frequently asked questions
What is an MCP server?
It is a software service that exposes tools, resources or prompt templates to a compatible AI application through Model Context Protocol. The server implementation and user permissions determine access to your systems.
How is MCP different from a regular API?
An API defines the operations of a specific system. MCP provides a shared way for AI clients to discover and call capabilities. An MCP server often uses existing APIs rather than replacing their business logic.
Can I connect ChatGPT to my own CRM?
Yes, through a custom MCP server when the required API and MCP connection support are available in your ChatGPT environment. Availability depends on the mode, workspace permissions and security settings. I verify this before development.
How do you connect Claude to a database?
Through an MCP server with focused read tools, restricted database views and permission checks. I verify transport and authentication support in the chosen Claude client. Unrestricted SQL access is unnecessary.
Does MCP work with local LLMs?
Yes, when the application or agent framework around the model includes an MCP client. The model itself does not automatically speak MCP. Tool calling, transport and result handling must be checked.
Is it safe to connect AI to company data?
Security depends on access policies, isolation, argument validation and prompt injection defences. I use least privilege, audit trails and human approval for sensitive operations. Permitted data and residual risks are agreed before launch.
How much does MCP server development cost?
There is no single fixed price. I estimate the work after reviewing APIs, workflows and access requirements. Hosting, third-party API and LLM usage costs are considered separately.
How long does an integration take?
Timing depends on API readiness, tool count, authentication and testing. Discovery produces milestones and a schedule. Unknown dependencies may require a technical feasibility check for one workflow first.
Can one integration connect several systems?
Yes. One server can expose tools for several systems, or a client can connect to multiple servers. The choice depends on access boundaries, ownership and isolation requirements.
Can the MCP server run on our own infrastructure?
Yes, on your VPS, in Docker or inside a private environment. Connectivity depends on the client. A remote client needs an accessible HTTPS endpoint or a supported secure channel; a local process can use stdio.
Related case studies and reading
FIGHTEX: economy and tokenomics simulatorAn example of game economy modelling and scenario testing.
MCP use cases for a websiteExamples of AI working with website tools and data.
MCP and RAG for working with business dataConnecting data retrieval with tools and result presentation.
Second Brain: knowledge base and vector searchExplore the project problem, solution and limitations.
MCP / AI INTEGRATION
Connect AI to your business systems
Tell me which systems you use and the work you want to automate. I will propose an MCP integration architecture and estimate the development scope.
@ifwcomMessage on TelegramDo not send passwords, API keys or confidential records.
Tell me about your project
Do not send passwords, API keys or confidential records.