Cohere study finds 2.6% of public MCP tools fully automate recognized work tasks

Cohere researchers released the Agentic Task Ecosystem (ATE) dataset on September 3, 2026, alongside an analysis of public Model Context Protocol tool listings reporting that 2.6% of the tools they matched to a recognized work task passed their strict test for performing that task end to end.

The released dataset contains 696,291 extracted tool rows drawn from 123,069 deduplicated public MCP server listings, collected across seven directories in May 2026. Within it, 18,058 of 694,411 rank-one tool-to-task matches were labeled good matches — the ratio that rounds to the headline 2.6%. That analysis-set denominator is smaller than the full 696,291-row tools table.

The Model Context Protocol is an open interface standard that lets an AI model call external software through a server, with each server advertising a list of named tools it can execute. Public directories catalog those servers, and Cohere collected from the directories themselves. The unit of analysis is a tool listing, not a running deployment or a recorded use.

To produce the 2.6% figure, Cohere used text embeddings to pair each analysis-set tool with its nearest software-performable task statement in O*NET, the standardized catalog of occupations and their component work tasks. A language model then judged whether the tool genuinely performed the named task rather than merely supplying information or completing a single step of it. The threshold is deliberately narrow, and Cohere states that the result does not measure tool adoption, workplace reliability, employment effects, or the share of jobs already automated.

Coverage was uneven across occupations. Cohere reports that matched tools reached 1,380 ONET task statements, and that 419 of 923 ONET occupations showed no agentic-tool activity at all in its analysis.

Those patterns only partly track earlier predictions of which jobs are exposed to language models. Cohere reports a 0.54 correlation across 178 occupations between realized MCP coverage and theoretical exposure scores drawn from the 2023 occupational-exposure study by Eloundou and co-authors, and no detectable relationship between those scores and whether tools reached an occupation’s routine or specialized work. That earlier study estimated potential exposure rather than observed deployment: it found that about 80% of U.S. workers could have at least 10% of their tasks affected and about 19% could have at least half affected, and its authors explicitly declined to predict development or adoption timing.

“ATE is a measure of supply, not adoption nor economic impact,” the Cohere Labs research team writes.

Most tools did not match a task at all. Cohere sorted the unmatched remainder into work narrower than an O*NET task statement, multi-task workflows, agent infrastructure, and a small set of apparently new work, and says it checked that classification against a blind human-labeled sample of 120 categories.

The corpus is bounded by what is publicly listed. It includes vendor-published enterprise servers that appear in public registries but excludes bespoke internal company deployments, and incomplete server documentation can cause tools to be missed. DataPhoenix has covered a different kind of enterprise agent deployment, Google Gemini Enterprise for Legal, a centrally managed product that sits outside this study unless its MCP tools appear in the public directories. Cohere characterizes 2.6% as a floor on that basis, though the released materials do not quantify the missing private tier or independently validate that every listed tool executes as described.

A separate analysis posted to arXiv in March 2026 examined 177,436 public MCP tools and found a comparable concentration in software work, reporting that software development accounted for 67% of tools and 90% of measured MCP server downloads. That study used a different corpus and classification method from ATE, so its shares are not directly comparable with the 2.6% figure.

No full technical report accompanied the release. The dataset card said a paper was coming soon; none was publicly available when the materials were checked on September 3, 2026.