- Airbyte's Pedro Lopez said MCP clients support different parts of the spec, so his team sometimes drops the spec feature and builds its own.
- Providers also add limits outside the spec, which shapes what builders can ship to agents.
Pedro Lopez, a software engineer at Airbyte, argued that MCP is applied unevenly. MCP is the Model Context Protocol, a shared standard for connecting AI agents to tools and data. He made the point on the AI Engineer show, in a talk about building an MCP server and a command-line tool for agents. Builders, he said, end up writing around the standard.
What was said
Airbyte built both interfaces so agents can reach company data in tools such as Zendesk, Stripe and HubSpot. Lopez said the MCP side keeps running into client differences. In his words, "even though there is a spec, the features that are supported across the clients kind of vary quite a bit."
His main example was elicitation, the part of MCP that lets a server ask the user for input. The first version, form mode, was not suited to sensitive data such as API credentials. A newer URL mode fits what Airbyte already does: send the user to a link to connect an account. But Claude Desktop told him it did not support URL mode. So the team went back to its own method instead of using the spec.
Sign-in showed the same pattern. The spec leans heavily on OAuth, and the main clients follow. Lopez said Anthropic's custom connector registration offered no place for a bearer token. His summary was blunt: "OAuth is the way."
He also named limits that providers add outside the spec. Claude caps the length of tool descriptions and limits how long tools can run. That makes long-running work harder over MCP.
Why it matters
Our reading: a vendor claiming MCP support is not promising that a tool works in every client. A team that ships for agents should test each client it cares about. It should also keep a fallback for features that some clients lack.
Buyers can ask a simple question of vendors: which clients have you tested, and what do you do when one lacks a feature? Lopez's examples suggest the answers will differ. Choices like OAuth-only sign-in are set by the clients, not by the builder.
The other side
Lopez did not call MCP broken. He said both MCP and the command-line tool have a place. MCP suits non-technical users, quick prototypes and one-off reports, and it lets tools in its ecosystem work with each other.
He also noted URL mode is newer than form mode. His examples are one company's experience with a few clients, and he did not say how many clients lag. He did not claim that this slows the wider web's move toward machines; that link is ours to test.
The command-line route has costs too. Packages are harder to keep up to date. And with no limits at all, he said, an agent can get stuck in loops.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Designing CLIs for Agents, Not Humans — Pedro Lopez, Airbyte (2026-10-09)
