Skip to content

AI, decoded

Should you use MCP or build a custom integration to connect AI to your tools?

Use MCP when a tool or data source will be reused across multiple models, apps, or teams — you integrate it once and everything speaks to it. Build a custom integration when you need a tight, high-control connection to a single system and the standard overhead isn't worth it. For most organizations scaling agents, MCP wins because it kills the N-models-times-M-tools integration explosion; custom is the exception for bespoke, performance-critical paths. Block's Angie Jones told Chain of Thought that with data stores connected through MCP, employees can query them in natural language from their preferred MCP-compliant agent instead of waiting maybe a week for a report.

· Updated · Chain of Thought

Level 4: How AI systems are built · 4.3 Tools, MCP and agent architecture

MCP (Model Context Protocol)AI Agents

  • 5 episodes
  • 5 guest clips
  • 3 glossary terms
A tangle of point-to-point wires beside a single hub, with the words: N times M wires, or N plus M. Wire each tool once with MCP.

In detail

Both get an AI model talking to your systems — a database, an API, an internal service. The difference is whether you build that connection once as a reusable standard or wire it bespoke each time. It’s the classic standard-vs-custom trade, applied to tool use.

1. Custom integration: maximum control, no reuse

A custom integration is code you write to connect one model to one tool. You control everything — auth, payload shape, error handling, performance — and for a single tight path that’s an advantage. The cost shows up at scale: every new model or app needs its own wiring to the same tool, and you maintain all of it. That’s the N×M integration explosion.

2. MCP: integrate once, reuse everywhere

The Model Context Protocol standardizes the connection. A tool exposes itself as an MCP server once, and any MCP-aware client can use it — no per-model glue. Block’s rollout of agents to 12,000 employees, which Angie Jones walked through on the show, leaned on exactly this: MCP turned tool access into something shared across the org instead of re-integrated per team. WisdomAI co-founder and CPO Kapil Chhabra says its hosted MCP server returns text and HTML that clients such as ChatGPT and Claude can display as interactive results.

3. MCP is an interconnect, not magic

Airbyte’s Michel Tricot put the boundary well on the show: “MCP is not dead. MCP is just an interconnect. It’s just a way of connecting an LLM to an external system.” A thin MCP wrapper over a weak API is still weak. MCP standardizes how you connect; it doesn’t fix what’s on the other end, and a bespoke integration can still beat it on a performance-critical path.

How to choose

Reused across models/apps/teams, or you want the broader MCP ecosystem → MCP. A single, bespoke, performance- or security-critical path → custom. Many orgs do both: MCP as the default connective tissue, custom integrations for the few places that earn the extra control. Yutori’s Dhruv Batra says that when sites lack APIs or tool interfaces, browser agents can find shipping costs by adding products to carts and entering ZIP codes across hundreds of e-commerce sites.

Why it matters

The integration layer is where agent programs stall — not the model. Defaulting to custom everywhere buries teams in glue code; defaulting to MCP turns tools into shared infrastructure, which is why it spread fast once agents went from demo to deployment.

Custom integrations multiply; MCP adds Left, custom integration: three models or apps each wired directly to four tools, a database, an API, an internal service and a data store, which makes three times four, twelve integrations to build and maintain. Right, MCP: the same three models and four tools, but each tool is exposed once as an MCP server and each model is an MCP client, so three plus four, seven pieces to build. Below, MCP is an interconnect, not magic: a thin layer in front of an API is still weak, and a custom integration still earns its place on a single performance- or security-critical path. Airbyte's Michel Tricot, episode 52: MCP is just an interconnect, a way of connecting an LLM to an external system. Wire it per model, or once for everyone Custom integrations multiply with every model and tool you add. A standard interface adds instead. CUSTOM INTEGRATION Model A App B Model C Database API Internal service Data store 3 × 4 = 12 integrations Each one built, secured and maintained on its own. MCP one protocol Model Aclient App Bclient Model Cclient Databaseserver APIserver Internal serviceserver Data storeserver 3 + 4 = 7 pieces to build Each tool is exposed once; any MCP-aware client can use it. An interconnect, not magic: a thin MCP layer over a weak API is still weak. “MCP is just an interconnect. It’s just a way of connecting an LLM to an external system.” Michel Tricot, Airbyte, episode 52. Custom still earns its place on a single performance- or security-critical path. Many orgs run both.
Custom wiring multiplies; a shared protocol adds. The counts are an illustration. The interconnect line is Airbyte’s Michel Tricot, from episode 52. Download the image

Hear it from the guest

“When people say MCP is dead, MCP is not dead. MCP is just an interconnect. It's just a way of connecting an LLM to an external system. You can make the LLM as complex as you want and as good as you want. Of course, if you just do a thin layer in front of an API, MCP is going to suck.”
Michel Tricot, Airbyte EP 52 · around 8:26 · read the transcript
Watch on YouTube
“Whereas now with MCP, we can connect to our data stores. And from your favorite … agent … that's MCP compliant, whether that's goose or something else.”
Angie Jones, Block EP 48 · around 22:26 · read the transcript
Watch on YouTube
“You don't want to work against one specific REST API of your data store or whatever other systems you have. But you will let your LLMs just figure it out”
Philipp Krenn, Elastic EP 32 · around 23:03 · read the transcript
“There is no API available for this. E-commerce websites expose that information after you add the item to cart, put in your zip code information, and then you'll see a listing of two or three options. Answering that question, requires a computer use or a browser use model”
Dhruv Batra, Yutori EP 63 · around 5:05 · read the transcript
Watch on YouTube
“Just connect the client where the work is happening with the wisdoms MCP server and the server. is also powered by mcp apps which essentially means that it is not just sending a text response back but it can also send a response in the html react format”
Kapil Chhabra, WisdomAI EP 75 · around 56:25 · read the transcript
Watch on YouTube

Quotes lightly edited to remove filler words.

Go deeper

  • MCP: Security best practices (2026-07-28) Model Context Protocol · Security guide (2026-07-28) Review the security work an MCP integration still needs, including consent, token handling and local-server risks.

From the conversation

This explainer is drawn from these episodes — each carries its full transcript.

Concepts in this explainer

Model Context Protocol (MCP)Tool Use (Function Calling)Large Language Model

Keep exploring