MCP – A bridge over the AI knowledge gap | karotammela.fi

For me, vibe coding was a fantastic way to start building with AI, but in my own projects I quickly ran into an invisible wall. It is the AI's limited awareness: a language model (LLM) is a sealed box whose worldview ends at the cutoff of its training data, often months or even more than a year before its public release. If you ask it about something that happened in the last year, the answer is mostly guesswork or outright hallucination. My solution has been MCP (Model Context Protocol). It is the bridge that connects the AI's "brain" to real-world data.

What is MCP?

Put simply, MCP is a standardised communication protocol that acts as a universal adapter between AI agents and external data sources. If the LLM is the computer's processor, MCP is its USB port. It lets the agent do more than just "talk": it can operate databases, call APIs, or use specialised tools that are not part of its core knowledge.

MCP as part of Kilo Code

I work with Kilo Code, which supports the MCP protocol natively. I don't even have to write the small glue snippet myself – I can install the most common MCP servers straight from the control panel. I just add a token and it's ready to go.

Configuration is made easy: In Kilo Code, MCP servers are defined in a kilo.jsonc file. You can place the configuration globally (~/.config/kilo/kilo.jsonc) or per project, with project-level settings overriding the globals.

Filling the one-year vacuum: Context7

For a long time, my biggest problem with agents was the "training data vacuum". The AI had no idea about the latest library updates or framework changes that had landed after its training cutoff. In my first green fee sales platform project this was by far the most time-consuming issue: many of Gemini's instructions or the dependencies it assumed were simply outdated. Copy-paste produced error after error, and debugging could drag on for a long time. I ended up fetching the latest documentation from the web and pasting it back into the conversation.

I solved this by wiring Context7 into my Kilo Code globally. It is a community-maintained, up-to-date code library. It gives the agent real-time context and documentation straight from the web. I don't start a project without it anymore – it is the difference between working code and "almost working" code. It fills the critical one-year gap that would otherwise make the agent unreliable.

The debate: Is MCP already dead?

Even though MCP works well most of the time and is only just over a year old, the developer community has been debating its future for a while now. Some see MCP as an unnecessary "extra layer" that adds complexity and bloats the context window.

CLI vs. MCP Many agentic developers now swear by raw CLI access (Command Line Interface). CLI's advantage is speed and flexibility: the agent can do anything on the command line without setting up a dedicated MCP server or negotiating a protocol.

MCP, on the other hand, represents a controlled and safer approach. It standardises what the agent is allowed to see and do. MCP's strength lies in its ecosystem – when someone builds a good MCP server, it is immediately available to every tool that supports the standard.

Critics still ask: why bother learning a new protocol if we can just grant the agent CLI access to the data? OpenClawn's developer Peter Steinberger, for example, swears by the CLI. On the Lex Fridman podcast Peter also digs into this topic.

Here's just one of the many articles written about it: MCP is dead, or MCP vs Skills revisited.

Summary

Whether MCP turns out to be the future standard or merely a stepping stone towards more autonomous CLI-driven agents, right now it is the most effective way for me to close the AI's awareness gap. It has sped up my own projects and given me the confidence that the code works on the first try, without version surprises popping up when I run the first build.