The Argument About MCP Is Ending; Its Adoption Is Not

The crest is behind it

For much of the past year the Model Context Protocol has been one of the loudest subjects in developer discourse. That noise has now begun to subside. Measured as a share of all tracked-entity stories on Hacker News and Lobsters, MCP's conversation peaked in March 2026 at 9.01% and has fallen in three of the four months since, settling at 6.62% in a partial June.

The decline in volume is matched by a decline in intensity. Average peak engagement per MCP story — the score a story reaches on the boards — moved from 11.4 in February to a high of 11.8 in March, then thinned to 7.4 in April and 6.0 in the partial June reading. The trajectory is uneven month to month, but the direction is consistent: fewer stories, and fewer people piling onto each one.

This is a cooling, not a collapse. May still carried 201 distinct MCP stories. The point is not that MCP is fading from view; it is that the marginal argument about it has been had. June is a partial month and should not be read as a finished data point, but it does not contradict the shape of the prior four.

A very large base to cool from

The cooling matters only in proportion to the height it falls from, and that height is considerable. In May 2026, MCP was the second most-discussed tool in our tracked set, with 201 stories. It sat ahead of Kubernetes (70), GitHub Copilot (66), React (52), Docker (43) and Cursor (37) — household names of the modern toolchain — and behind only Claude Code, at 349.

That ordering is the corrective to any narrative of decline. A subject that out-discusses Kubernetes and Docker combined, in the month its own conversation is supposedly fading, is not fading into irrelevance. It is settling onto a plateau that most tools never reach. The persistence of Claude Code at the top also extends a thread from our earlier coverage of the AI coding contest: the assistant leads the conversation, and the protocol it speaks sits directly behind it.

The talk cools; the installs climb

The reason to call this maturity rather than decay lies in the adoption data. Across the snapshot window of June 12–19, MCP's weekly npm downloads rose from 38.36 million to 39.84 million, peaking at 39.84 million on June 18. Over the same window, weekly PyPI downloads rose from 50.0 million to 65.0 million, and GitHub showed roughly 1,439 new MCP-related repositories in a single weekly reading.

These figures carry real caveats. The window is about a week — a snapshot, not a trend — and the PyPI increase is concentrated in the final reading, so it should be read as rising rather than annualized into a growth rate. The repository count is a single reading on June 15. With those qualifications stated plainly, the divergence still holds: the lines of conversation and consumption are pointing in opposite directions, which is precisely what one expects as a technology stops being argued about and starts being depended upon.

This is also a distinct angle from our earlier work. We have previously framed MCP through the lens of the toolchain as an attack surface. This issue is the complementary maturity story: we called it a target; now we watch it become furniture — the thing people stand on without remarking on it.

The pattern is MCP's, not the category's

One might object that the whole agent-infrastructure layer is simply quiet, and MCP is merely sharing in a general lull. The data refutes that. Across February to June, the rest of the category sat in the low single-to-mid double digits per month: LangChain ran 16, 15, 13, 13 and 5 stories; Ollama 17, 37, 28, 11 and 16; vLLM 10, 14, 9, 11 and 13; n8n and crewai in the low single digits.

Against those numbers MCP's run of 315, 378, 272, 201 and 145 stories is not category drift — it is one protocol carrying the entire conversation almost single-handedly. The divergence between talk and adoption is therefore specific to MCP, and reads as the maturation of a particular standard rather than a fashion passing through a sector.