Cohere ATE Maps Nearly 700,000 MCP Tools to Occupational Tasks, Finds Only 2.6% Can Complete Them End to End

Cohere Labs released the Agentic Task Ecosystem (ATE) on September 3, organizing 696,291 tools and 123,069 MCP servers collected from seven public directories in May 2026 into downloadable Parquet data. The team first deduplicated entries across directories, then identified the closest U.S. O*NET task description for each tool and asked a language model to classify the match as end-to-end execution, support for only part of the workflow, or an incorrect match. The dataset also retains tool names and descriptions, occupations and SOC codes, cosine similarity scores, classification rationales, and quality labels, allowing researchers to apply their own filters.

Under its strict definition, only 2.6% of tools were judged capable of completing an O*NET task from start to finish. Tools that merely retrieve data, send messages, or handle a single step do not count as full automation. Successful matches covered 1,380 distinct task descriptions, representing about 15% of tasks considered executable by software. Of 923 occupations, 419 had no matching tools at all. Across 178 occupations, theoretical AI exposure and actual MCP coverage had a correlation of 0.54, but tool counts alone cannot reveal whether those tools address routine work or an occupation’s core professional responsibilities.

The dataset is significant because it provides a queryable snapshot of the supply side of agent tools, enabling research into duplicated development, capability gaps, and occupational distribution across the MCP ecosystem. However, it includes no data on downloads, real-world deployments, success rates, security, or long-horizon task testing, and tool descriptions may overstate actual implementations. Nearest-neighbor retrieval followed by language-model classification can still produce tenuous matches, while human validation covered only some categories. For engineering and research teams, the next step should be to add executable MCP conformance tests, versioning and usage telemetry, and occupational-task ontologies for Chinese and other regional contexts. Until then, the 2.6% figure should be interpreted only as a supply-side indicator under a particular classification scheme—not as a measure of how much work agents can already replace.