Vibe coding is more than a development method. By opening software creation to more people, it could affect businesses of every size and industry. Understanding it helps companies keep pace with technology and guide clients through AI-assisted development.

Here, we explore what vibe coding means, its advantages and limitations, when it makes sense and how it compares with traditional programming.

What is vibe coding?

Definition and origins

Vibe coding uses AI to turn natural language prompts into functional applications instead of relying on manual programming. OpenAI co-founder and former Tesla AI leader Andrej Karpathy coined the term in February 2025.

Karpathy described it as “fully giving in to the vibes, embracing exponentials, and forgetting that the code even exists”. It replaces the precise, methodical emphasis of traditional programming with an intuitive, experimental approach powered by large language models (LLMs).

The idea follows Karpathy’s 2023 observation that “the hottest new programming language is English,” reflecting how AI has made natural language a practical programming interface. By March 2025, Merriam-Webster had listed vibe coding as a “slang & trending” term.

What are the key characteristics of vibe coding?

Four characteristics distinguish the approach:

• Natural language interface: users describe requirements and functionality in plain English instead of programming syntax. This opens development to non-technical users and can accelerate experienced developers’ work.
• AI-driven code generation: LLMs interpret descriptions and produce functional code, taking care of implementation details from basic algorithms to complex application structures without manual intervention.
• Iterative refinement: the emphasis is on “code first, refine later”. Users test generated solutions, then supply further prompts to improve functionality rather than aiming for perfect code immediately.
• Minimal code review: accepting generated code without comprehensive human review is a defining feature. As programmer Simon Willison noted, true vibe coding means using code “without full understanding” of how it works.

How vibe coding works

The process

Vibe coding moves from an idea to a working application through four steps.

Step 1: Choose an AI coding platform. Match the assistant to your technical requirements, performance needs and budget. Options include Replit, Cursor and cloud-based development environments.

Step 2: Define requirements. Describe the intended functionality clearly and specifically. Prompt quality directly influences output quality. Instead of “create an app,” try: “Create a startup name generator app with a text input for industry selection and a button that displays 10 creative company names when clicked”.

Step 3: Refine the code. Review the initial output, test the application and identify improvements. Use conversational feedback to fix issues, adjust functionality or add features.

Step 4: Review and deploy. Assess the refined code before deployment. Although less comprehensive than a traditional development review, this final check establishes whether the application meets basic functional requirements.

Popular tools and platforms

Several platforms support AI-assisted development, each with a different emphasis.

Google’s development platforms cover the path from idea to application. Google Cloud AI Studio offers a fast route, often starting from one prompt, with live previews and direct deployment to Cloud Run for public access.

Firebase Studio complements this with backend capabilities such as authentication and databases for production-ready applications. It creates detailed app blueprints for review before generating working prototypes.

Cursor is an AI-native editor built on VS Code. It integrates models such as Claude Sonnet for writing, editing and understanding code. Composer lets users delegate tasks to AI agents with context across multiple files.

GitHub Copilot led the cited survey, with 42% of engineers naming it their primary AI coding tool. Microsoft’s product supplies code suggestions within existing development workflows.

Replit offers a browser-based development environment with AI assistance through Ghostwriter. It supports rapid prototyping and collaboration without requiring a local setup.

Claude Code uses Anthropic’s models for code generation, debugging and natural language programming. It supports complex requirements and the creation of clean, documented code across multiple languages.

Windsurf is presented as a cloud-native IDE for vibe coding, combining real-time collaboration, AI assistance and deployment options. It pairs familiar development workflows with AI capabilities for rapid application creation and testing.

What are the benefits of vibe coding?

Speed and efficiency

Vibe coding can compress development from weeks or months into hours or days. The cited research reports up to 55% faster project completion compared with manual coding.

This is particularly valuable for prototypes and minimum viable products (MVPs). Businesses can validate ideas, test market demand and refine concepts without committing substantial development resources upfront. Moving from an idea to a working prototype within hours supports more agile strategies and faster responses to market opportunities.

Innovation and creativity

Automating routine coding gives developers more time for high-level problem-solving and creative design. Teams can concentrate on user experience, business logic and distinctive features instead of syntax and implementation details.

This supports what industry experts call a “problem-first approach”: prioritizing user needs over technical constraints. Teams can quickly test several possible solutions without the time normally needed to implement each manually.

The benefits also reach beyond developers. Non-technical team members can contribute directly, bringing domain knowledge and user insights that might otherwise be lost when business requirements are translated into code.

Accessibility and democratization

People without programming backgrounds can describe what they need and create functional applications. That substantially lowers the barrier to software development.

Domain experts can build solutions to industry-specific problems without depending entirely on technical intermediaries. Small businesses and startups can access custom software without the traditional cost of hiring specialized developers.

What are the risks and limitations of vibe coding?

Security vulnerabilities

AI-generated code can contain security flaws and outdated practices that expose production systems. The cited research reports basic vulnerabilities in 9 out of 10 AI-generated projects, which non-technical users may be unable to recognize or resolve.

Speed can amplify the problem when teams rush code into production without adequate review. Traditional development incorporates security into implementation; a vibe-coded “black box” can hide vulnerabilities until they are exploited.

Organizations need robust scanning processes and appropriate security review before deployment. DataNorth AI’s EU AI Act compliance service helps businesses address the regulatory and security implications of AI-assisted development.

Technical debt and maintainability

Generated solutions may depart from coding standards and architectural best practices, quickly accumulating technical debt. Applications “stitched together by vibes alone” can lack the structure and documentation required for long-term maintenance.

Debugging becomes harder when developers do not fully understand the code’s logic or implementation. That gap matters when updating applications, fixing bugs or integrating other systems. AI can also generate inconsistent coding patterns, adding to the maintenance burden.

Scaling introduces further challenges. Tools that work well for small or medium projects may struggle with distributed applications and complex architecture. Growing a vibe-coded application can therefore require substantial refactoring or a rebuild.

Quality and reliability concerns

AI-generated code can contain functional bugs, logic errors and performance problems that initial testing misses. Conversational refinement does not guarantee comprehensive quality assurance or systematic test coverage.

The cited study reports that 25% of developers estimate one in five AI-generated suggestions contains factual errors or misleading code. Such risks require additional quality controls in production or when applications handle sensitive data.

Models also have limitations with unfamiliar or complex problems. They handle standard algorithms and common functionality more readily than unique requirements or emerging technical challenges that demand original problem-solving.

Vibe coding vs traditional programming

These differences help businesses decide which development approach fits their needs.

When to use vibe coding

Vibe coding is most useful when speed and accessibility matter more than code complexity or long-term maintenance.

Rapid prototypes and proofs of concept are its strongest use case. Working prototypes built in hours rather than weeks help validate business ideas and demonstrate functionality to stakeholders. This supports iterative design and faster market validation.

Internal tools and automation are another opportunity. Teams can build dashboards, data processing scripts and workflow tools without major development investment. The source describes these applications as typically having lower security and scalability requirements, making them suitable candidates.

Educational and experimental projects let teams explore new technology and ideas. The low entry barrier makes experiments possible that traditional development might not justify.

Startup MVPs have also embraced AI-assisted development. Reported data indicates that 25% of Y Combinator’s Winter 2025 startups built 95% of their codebase with AI assistance, illustrating its appeal for companies seeking rapid market entry.

When not to use vibe coding

Some applications need the precision and control of traditional programming.

Production systems handling sensitive data require comprehensive security review and custom protections that vibe coding typically cannot provide. Finance, healthcare and other regulated sectors need detailed security control.

Large, long-term applications benefit from deliberate architecture and structured development. Systems with many external integrations, demanding performance requirements or extensive customization exceed the capabilities of most vibe coding tools.

Mission-critical applications should avoid vibe coding when reliability is essential. Software whose failure could cause safety risks, financial losses or operational disruption needs comprehensive testing and validation.

Best practices for safe vibe coding

Security guidelines

Organizations need a clear security framework for AI-assisted development.

Require human review of all generated code before production deployment. Even when teams embrace the vibe coding philosophy, critical applications need security expertise to identify vulnerabilities introduced by AI.

Supplement that review with automated scanning. Static analysis security testing (SAST) and dynamic analysis security testing (DAST) can detect common vulnerability patterns and provide another layer of protection.

Define where vibe coding is appropriate and where traditional development is required. Clear boundaries ensure higher-risk systems receive suitable oversight.

Keep experimentation separate from production. Isolated environments prevent experimental applications from affecting live systems or sensitive data and reduce the risk of accidental breaches.

Quality assurance

Rapid development still requires systematic quality practices.

Combine vibe coding with established testing methods, including unit tests, integration tests and user acceptance testing. Quickly generated functional code still needs comprehensive validation.

Make sure team members understand critical components. Production work requires knowledge of at least the overall architecture and essential business logic, even if the underlying philosophy accepts code without full comprehension.

Document the prompts, intended functionality and changes made during refinement. These records support future maintenance and troubleshooting.

Apply version control and change management to generated code as you would to manually written code. That includes branching strategies, code reviews and deployment procedures.

The future of software development

Industry trends

AI-assisted coding is reshaping software development. The cited forecasts for 2030 predict that 25% of Y Combinator startups will generate most of their code with AI, routine tasks will become 80% automated and project completion will accelerate by up to 55%.

AI-generated pipelines and autonomous software development lifecycle loops are emerging as the next stage. These systems are expected to cover requirements, deployment and monitoring, allowing developers to concentrate on strategy, ethics and oversight.

Boundaries between low-code, no-code and professional programming are also narrowing. Platforms increasingly combine visual tools, natural language prompts and traditional code within interfaces that accommodate different skills and working methods.

Enterprise adoption is growing too. The cited 2025 reporting puts agentic AI usage at 82% of companies for software development tasks, compared with 50% in December 2024, pointing to a substantial change in development practices.

Impact on developers

Developers’ roles are shifting from implementing every line to coordinating AI and designing systems. Defining requirements, reviewing generated output and maintaining security and quality standards become increasingly central.

Prompt engineering, AI tool management and solution architecture are becoming more important skills. Programming knowledge remains valuable, while effective collaboration with AI increasingly matters for career progression.

Hiring practices reflect that shift. Companies including Visa, DoorDash and Roblox have included “AI coding fluency” in job descriptions and specified AI use for 50% or more of engineers’ workflows.

Development environments are also becoming multimodal, combining voice, visual and text interfaces. These advances promise greater accessibility and more intuitive workflows while improving productivity for experienced developers.

Getting started with vibe coding

Implementation steps

A structured rollout helps organizations explore the approach.

1. Begin with low-risk experiments. Internal tools, prototypes and proofs of concept help teams understand the tools’ strengths and limitations while limiting potential business impact.
2. Choose tools around your team and projects. Consider skill levels, requirements and budgets. The cited adoption figures put GitHub Copilot at 42%, while Cursor and Replit offer different advantages for particular uses.
3. Set security and quality rules before expanding. Define review processes, scanning requirements and documentation standards that fit your risk tolerance and compliance obligations.
4. Train the team. Cover prompt engineering, tool usage and quality assurance. Teams need to communicate effectively with AI and evaluate its output.
5. Expand gradually using evidence. Track development speed, code quality and security outcomes, then use those results and lessons from initial projects to guide wider adoption.

Conclusion

Vibe coding is changing software development by making application creation more accessible and accelerating experimentation. Prototypes, internal tools and MVPs can help businesses validate ideas and respond to opportunities faster.

Those benefits come with security, quality and maintenance challenges. Organizations that combine appropriate guardrails, training and expert guidance can gain advantages while avoiding common mistakes.

The likely direction is a hybrid approach: the speed and accessibility of vibe coding alongside traditional programming’s precision and control. Understanding when to use each supports sustainable growth and innovation.

DataNorth AI helps organizations make this transition through expertise, training and strategic guidance, while maintaining risk management and quality standards. Explore our Vibe Coding Workshop to equip your team with the tools and guidelines it needs.

The shift is already underway. The challenge is how quickly and effectively your organization can adapt.

Originally published by Jorick van Weelie at DataNorth, updated 17 December 2025.

Large Language Models (LLMs), familiar through applications such as ChatGPT, often attract attention for their broad capabilities. Businesses are also exploring a more focused and often more practical option: Small Language Models (SLMs). These offer an accessible, efficient way to bring AI into core operations and tailor it to specific needs.

This article explains what SLMs are and how they differ from larger models. It explores their strategic advantages, practical applications and the considerations that help businesses put them to work.

Understanding Small Language Models

Small Language Models take a focused approach to AI, with streamlined architectures and performance optimized for specific tasks. Their parameter counts typically range from millions to a few billion, often staying below 10 billion.

Designed for efficiency, they can operate on devices with limited resources or in settings where connectivity, memory and electricity are constrained. SLMs are frequently fine-tuned on smaller datasets from a particular domain, creating specialized tools for specific uses.

Small Language Models vs LLMs

The differences between small and large language models shape both their capabilities and how businesses deploy them.

Size and model complexity

Large models such as GPT-4 have been estimated to contain as many as 1.76 trillion parameters. GPT-4 is also speculated to use a Mixture of Experts (MoE) architecture. By comparison, the open-source Mistral 7B has around 7.3 billion parameters.

Scale affects more than model size: it influences computing requirements and the range of tasks a model can handle.

MoE models are sparse, activating only a subset of their parameters for a given input. Even with an estimated total of 1.76 trillion parameters, a single forward pass would use far fewer. MoE architectures also support learning across a broad range of tasks, while dense variants are more rigid.

Training and resources

Training LLMs requires enormous datasets and computing capacity, often involving thousands of GPUs. This brings substantial infrastructure costs and continuing operational expenses.

SLMs require considerably less computing power, memory and storage, making them more affordable to train and maintain.

Versatility vs. specialization

LLMs are built for broad, general-purpose work, seeking to emulate human intelligence across many domains. Their strength is the variety of applications they can support.

SLMs focus on precision and efficiency within particular domains. This makes them useful for tasks where targeted accuracy matters most.

Inference speed

With fewer parameters, SLMs can generate responses faster and with lower latency. LLMs can respond more slowly, particularly when many users access them simultaneously.

Bias

SLMs carry a lower risk of bias because they are trained on smaller, domain-specific datasets. LLMs often draw on vast amounts of unfiltered internet data, where misrepresentation and under-representation can introduce bias.

Deployment

SLMs are easier to deploy on local machines, edge devices and mobile applications, often without substantial server hardware. LLMs typically require powerful cloud servers to operate.

Applications for Small Language Models

SLMs are finding practical uses in specialized industry settings.

Healthcare

SLMs can be customized to analyse medical text and provide real-time health monitoring and advice through smart wearables. Operating without a continuous cloud connection is important for maintaining patient privacy.

Finance

These models can analyse earnings calls and financial news to inform stock predictions. They can also assess loan documentation for credit scoring and identify changes in sentiment within market commentary.

Agriculture

SLMs are being developed to answer complex agronomic questions and help farmers make decisions in real time.

Legal

SLMs can examine contracts to identify and categorize clauses, flag potential risks, and extract important dates and financial terms for compliance tracking.

Marketing

SLMs can help produce marketing materials, including social media posts and product descriptions.

Your next step

Small Language Models offer an accessible, cost-effective route to integrating AI into core business operations. They bring precision to specific tasks, enhance data privacy and support real-time applications that can change how organizations operate and serve customers.

Successful implementation starts with understanding your business needs and choosing suitable tools. Whether you want to automate customer support, streamline data processing or develop on-device AI, SLMs offer a practical opportunity. Model selection, fine-tuning and deployment at scale can still be challenging.

DataNorth’s AI experts can help align your AI initiatives with business objectives and maximize your return on investment.

Originally published by Jorick van Weelie at DataNorth, updated 25 July 2025.

Building an LLM application might sound like a project reserved for large organizations with extensive resources. Training a model from scratch does demand substantial resources, but building an application that uses an LLM is now much more accessible to beginners.

Even for most complex applications, you don’t need to train a model from the ground up. Using a pre-trained model, or fine-tuning one for your needs, is often a more practical and resource-efficient approach.

Step 1: Define your goals and use case

Start by deciding what you want your LLM to do. What purpose will your application serve? Which complex tasks should the model help with?

These goals will guide later decisions, including which LLM to choose and what high-quality data you might need.

Step 2: Choose an LLM

Once you’ve defined your application’s purpose, you have three main options for powering it:

• Use a pre-trained model: This is the simplest and most cost-effective option. You use a model already trained on a large, general dataset for inference, or making predictions, without further training. Because no additional learning takes place, this approach requires the least computation and data.
• Fine-tune a pre-trained model: Start with an existing model and train it further on a smaller dataset relevant to your use case. Fine-tuning helps the model become highly proficient at a specialized task or within a particular domain. It requires more computation than using a model unchanged, but significantly less than training from scratch.
• Train a model from scratch: This is generally not recommended for most applications. Building an LLM from the ground up begins with randomly initialized weights and requires enormous amounts of data and computing power, with many potential pitfalls. Pre-trained models are almost always the starting point for fine-tuning because their foundation of general knowledge saves substantial resources and effort.

Note: Platforms such as Hugging Face offer a wide range of open source pre-trained models and fine-tuning tools. Libraries such as PyTorch and TensorFlow provide extensive tools for working with these models.

Step 3: Prepare your data

Your choice of LLM strategy determines the data preparation you need.

For fine-tuning, gather a dataset specific to your task or domain. Its quality is crucial to performance. Preparation involves collecting relevant text, cleaning it and, where needed, formatting or annotating it.

When using a pre-trained model, you still need to consider the input your application sends it. Relevant, well-formatted data helps the transformer architecture process the information effectively.

Step 4: Integrate the LLM

With your model selected or prepared, connect it to your application. This often means setting up an API.

Many model providers offer APIs for accessing their LLMs. For a fine-tuned open source model, you would set up an API endpoint to serve it. With pre-trained models, the designated provider usually supplies the endpoint. Your application sends requests to the API, which returns the model’s response.

Step 5: Build the application interface

Next, create the user interface. Decide how people will submit input to the LLM and how they will see its output.

For a chatbot, for example, you need a text input field and an area displaying the conversation. Your application code takes the user’s input, sends it to the model through the API, receives the response and displays it.

Step 6: Evaluate and test

Once the application is assembled, test it thoroughly to check that the LLM performs as expected for your use case.

Use realistic inputs and assess whether the outputs are accurate, relevant and helpful. Applications for a particular domain may need evaluation criteria specific to that domain. You can use evaluation datasets or ask people to assess response quality, especially for open-ended tasks.

Step 7: Deploy and monitor

The final step is making your application available to users and keeping it working correctly over time.

Deployment involves hosting the application and LLM on servers, often through cloud platforms such as AWS, Google Cloud or Azure, which provide scalable resources.

Monitor the application continuously to track performance and model behaviour, identify performance issues and catch unexpected outputs. Periodically updating or retraining the model with new data can help keep it relevant and performing well.

Challenges when building your own LLM

Computational costs: API use and fine-tuning consume computing resources and incur costs, particularly at scale.

Data privacy: If your application handles sensitive data, both the application and the LLM must comply with privacy regulations.

Debugging: Getting the model to behave as intended can take experimentation, especially when fine-tuning or engineering prompts.

Conclusion

You don’t need a large team or an extensive AI background to build an LLM app today. Clear goals, quality data and the right tools can help you develop powerful applications using modern LLMs.

Ready to build your own private LLM, or looking for expert guidance through the process? Contact us today!

Originally published by Nick Moesker, DataNorth*, on* 30 June 2025*:* Building your own LLM application: A beginners guide.

What is tokenization?

Tokenization breaks human language into smaller, more manageable units called tokens.

Every input sent to an LLM passes through a tokenizer. It turns the input into a sequence of tokens, representing words with numbers because LLMs can only process numerical data.

The tokenization process

There are three core steps:

Input data: The raw text is handled before it reaches the model. The nature of that text determines how it should be divided into tokens.

Transformers: This is the model architecture that processes tokenized input. Tokenization puts the input into the right format so the model can process it effectively.

Prediction output: The model generates predictions, but the quality of its output is influenced by how well the input was tokenized. Good tokenization preserves the original text’s meaning and structure, helping the model produce more accurate, coherent responses.

Types of tokenization

Text can be tokenized in three main ways:

• Word based: Each word becomes an individual token.
• Character based: Each character becomes a token.
• Subword based: Words are divided into subword units rather than kept whole or split into individual characters.

What is a subword?

Subwords are smaller, meaningful units within a word.

For example: unstoppable → un + stoppable

This helps a model understand and generate variations of words it may never have encountered before. That ability is crucial during training, particularly with large, diverse datasets.

LLMs primarily use subword based tokenization because it balances vocabulary size with contextual understanding.

Disadvantages of word based tokenization

Repetition of similar words: LLMs struggle to connect words with the same root. For example, “run,” “running,” and “runner” are treated as separate tokens without recognizing their relationship.

Filler words: Common words such as “and,” “or,” “the,” and “is” occupy token space without adding significant meaning to a sentence.

Out-of-vocabulary issues: Models struggle with words they did not encounter during training, including spelling mistakes and new terminology.

Disadvantages of character based tokenization

Small vocabulary, long sequences: A small vocabulary comes with much longer token sequences, making processing inefficient.

Lack of meaning: Individual characters carry no semantic meaning on their own, making context and relationships harder for models to understand.

Loss of language structure: This approach misses important patterns and word structures needed for language understanding.

How do we decide where to split words?

Byte Pair Encoding (BPE) is a popular approach. It merges frequently occurring character pairs or subwords to reduce the overall token count while retaining meaning. Its efficiency makes it widely used in natural language processing systems.

Tokens and pricing

Using AI comes at a cost, at least when you need advanced features.

Tokenization directly affects LLM pricing. Rather than paying a fixed amount for every interaction, you pay according to the number of tokens you use.

API customers typically encounter two types of pricing:

• Input pricing: The cost of tokens in your prompts and queries.
• Output pricing: The cost of tokens in the model’s response.

Providers often list prices per 1,000 or 1 million tokens, helping you estimate the cost.

Token efficiency: How many tokens for the same sentence?

Consider the sentence: “Artificial intelligence is transforming the world.”

• GPT-4: 8 tokens
• Claude: 8 tokens (estimated)
• Mistral: 9 tokens
• LLaMA 3: 10 tokens

Pro tip: Tools such as ChatGPT’s tokenizer page show how many tokens your text contains and their corresponding token IDs.

Why tokenization matters

Anyone working with LLMs benefits from understanding tokenization because it affects:

• Cost management: Token counts directly influence API costs.
• Performance optimization: Efficient tokenization supports better model performance.
• Prompt engineering: Understanding how text is tokenized helps you write more effective prompts.

Tokenization may seem technical, but it is fundamental to understanding how LLMs interpret and process language. At DataNorth, we empower businesses to use and implement artificial intelligence in their operations.

Curious how AI can optimize and facilitate your processes? Get in touch with us to explore the possibilities!

Originally published by Nick Moesker, DataNorth*, on* 5 June 2025*:* Tokenization for LLMs.

The Model Context Protocol (MCP) is an open standard for connecting AI applications to external tools and data, including files, databases, business software and APIs. Anthropic introduced MCP in November 2024. It now sits within the Agentic AI Foundation under the Linux Foundation, with support from OpenAI, Google, Microsoft and AWS.

• A tool exposed through an MCP server works across AI applications that support MCP, removing the need for individual integrations.
• The current specification, dated 2026-07-28, made the protocol stateless, simplifying server scaling.
• MCP standardizes connections. You still decide which servers to trust.

What is the Model Context Protocol?

MCP provides a standard way for AI applications to discover and use external tools and data. The official documentation describes it as a USB-C port for AI applications: wrap a tool in an MCP server once, and any application that speaks the protocol can use it.

David Soria Parra and Justin Spahr-Summers created MCP at Anthropic, which released it as open source on 25 November 2024. In December 2025, Anthropic donated MCP to the Agentic AI Foundation. It co-founded the foundation with Block and OpenAI, supported by Google, Microsoft, AWS, Cloudflare and Bloomberg.

The name can be confusing. Hopsworks’ MLOps glossary uses “model context protocol” for a concept that records a model’s context, including references to training data, hyperparameters and environment, to make results reproducible. Anthropic’s MCP serves a different purpose: it connects systems and does not store datasets or training settings.

Key benefits of the Model Context Protocol

Putting one protocol between many AI applications and tools brings several benefits.

• Standardisation: ten applications and twenty tools require thirty MCP implementations, one for each, rather than two hundred custom integrations.
• Interoperability: Claude, ChatGPT, Microsoft Copilot Studio, the Gemini API, VS Code and Cursor support MCP. An integration can therefore be reused when switching applications or models.
• Composability: a task can draw on several servers, such as reading a Jira ticket, checking GitHub code and posting a summary in Slack.
• Current context: models can use live source-system data instead of relying solely on training data, a core aspect of context engineering.

MCP architecture: how it works

MCP follows a client-host-server architecture, allowing one AI application to connect to multiple servers. The specification defines three roles.

Tools are functions a model can call. Resources are data an application can read, while prompts are templates users can select. Messages use JSON-RPC 2.0, carried over stdio for local servers or Streamable HTTP for remote ones. Remote servers requiring authorization use OAuth 2.1.

The 2026-07-28 revision made MCP stateless. It removed the initialization handshake and protocol-level sessions; each request now includes its own protocol version and capabilities. Any server instance behind a load balancer can consequently handle any request.

The host uses one client per server. Each server wraps an existing system and exposes it as tools, resources and prompts.

Comparing data integration approaches

MCP usually wraps existing APIs rather than replacing them. The distinction is who builds the integration and how often it needs building.

An MCP workflow: scheduling a meeting

An AI assistant scheduling a meeting through a calendar MCP server follows this sequence:

1. The client requests the calendar server’s tools, receiving functions such as find_free_slots and create_event.
2. You request a one-hour meeting with three colleagues next week. The model calls find_free_slots.
3. The server queries the calendar API and returns available slots.
4. The model suggests a time. The host asks you to confirm before running create_event.

The server creates the invitation, and the assistant reports the outcome.

This pattern applies across MCP servers: the model selects a tool, the host controls its execution, and the server communicates with the underlying system.

Is MCP secure?

MCP can support secure integrations, but using the protocol does not make an integration safe by itself. The specification sets out security principles, including explicit user consent before hosts invoke tools. It also states that “MCP itself cannot enforce these security principles at the protocol level”.

Two incidents illustrate the risks. In May 2025, researchers found that a malicious GitHub issue could hijack an agent using the GitHub MCP server and leak private repository data. In September 2025, Postmark warned that a fake postmark-mcp package had secretly sent copies of emails to an outside server.

Reduce exposure by connecting only trusted servers, limiting each server’s permissions, requiring approval for tools that change data and logging every tool call. Our article on AI guardrails explores broader safeguards for AI applications.

MCP’s role in AI evolution

In under two years, MCP has moved from an Anthropic project to shared infrastructure. Token cost remains its main practical drawback.

• Adoption: OpenAI built its Apps SDK for ChatGPT on MCP. Microsoft supports it in Copilot Studio and VS Code. Google Cloud lists more than 50 Google-managed MCP servers, either generally available or in preview, while AWS’s own MCP server became generally available in May 2026. In July 2026, MCP maintainers reported “close to half-a-billion downloads a month” across the official Tier 1 SDKs.
• Token cost: each connected server adds tool descriptions to the model’s context, consuming tokens. Anthropic demonstrated that asking a model to write code to call MCP tools reduced one example workflow from 150,000 to 2,000 tokens.

Practical applications

MCP connects AI applications with existing systems in several settings.

• Software development: assistants including Claude Code, GitHub Copilot and Cursor use MCP servers to read issues, query databases and find documentation. They often combine these with skills explaining how to use the tools.
• Business assistants: ChatGPT apps, Copilot Studio agents and Claude connectors access CRM, email, calendars and documents through MCP.
• Internal systems: organizations build MCP servers for their ERP or CRM, making one integration available to every MCP-compatible AI application they use. GDPR applies whenever these systems contain personal data, just as with any integration.

Conclusion

The Model Context Protocol has become a shared standard for linking AI applications to tools and data. It cannot decide which servers you should trust. Give every MCP integration an owner, minimal permissions and logging, as you would any other integration.

Wondering which systems to connect to AI first and how to connect them safely? Our AI consultancy helps you choose and build the right integrations.

Frequently asked questions (FAQ)

What does MCP stand for in AI?

MCP means Model Context Protocol. Introduced by Anthropic in November 2024, this open standard defines how AI applications including Claude, ChatGPT and Microsoft Copilot Studio connect to external tools and data through MCP servers.

Who developed the Model Context Protocol?

David Soria Parra and Justin Spahr-Summers created MCP at Anthropic. It was released as open source on 25 November 2024 and has belonged to the Agentic AI Foundation under the Linux Foundation since December 2025.

What is an MCP server?

An MCP server exposes a tool or data source to AI applications through tools, resources and prompts, usually built on an existing API. It can run locally or remotely over HTTP.

What is the difference between MCP and an API?

MCP builds on APIs. An MCP server typically wraps an existing API and describes its functions in a standard format, allowing any MCP-compatible AI application to use it without a custom integration.

How does MCP relate to A2A and ACP?

MCP connects agents to tools and data. The Agent2Agent protocol (A2A) supports collaboration between agents in multi-agent systems. IBM’s Agent Communication Protocol (ACP) merged into A2A in August 2025. ACP may also mean Stripe and OpenAI’s Agentic Commerce Protocol or Zed’s Agent Client Protocol.

Originally published by Jorick van Weelie at DataNorth.