Cohere has published an open dataset mapping nearly 694,411 public MCP tools across 178 occupations, offering a supply-side view of which kinds of work developers are making available to AI agents. Its key finding is a mismatch: only 2.6% of tools cleanly match a complete recorded occupational task.
Cohere Watch analysis
What happened
The Agentic Task Ecosystem dataset draws on 123,969 public MCP server listings. Cohere’s researchers compare tool descriptions with ONET task statements, then use a broader taxonomy to capture work that does not fit those recorded tasks neatly.
The researchers say tool coverage varies by occupation. Tools reach relatively specialised tasks in computing and legal work, while coverage in production and healthcare leans towards more routine tasks. They also find that tools reach further into specialised work when handling digital information is central to an occupation’s expertise.
Why it matters
A list of available tools is not a tally of jobs already automated. Cohere’s strict matching finds that tool descriptions and recorded work tasks often do not line up: tools can cover smaller components of a task, or bundle several tasks together. The 2.6% figure therefore describes a narrow match between two kinds of record, not the share of work that agents can or cannot do.
That distinction matters for anyone trying to read the labour-market tea leaves from agent directories. MCP tools show what developers have made available, not how often people use them, whether they work reliably, or what happens inside private company systems.
Our read
This is a useful map of the agent ecosystem, not a jobs forecast. Its value is in showing where public tools are being built and how awkwardly that supply fits existing occupational categories. The dataset is open, which gives other researchers room to test the picture rather than merely admire it from the vendor’s foyer.
What to watch
- Whether later research connects public tool availability with evidence of actual use.
- How tools chain together into longer workflows, a gap the researchers identify.
- Whether private, bespoke company workflows change the occupational patterns in this public dataset.
- How future work represents newer or less clearly recorded kinds of work.
Discussion spark: When judging AI’s effect on work, should researchers start with the tools developers have made available, or wait for evidence of how people actually use them?
Sources and evidence
- The Agentic Task Ecosystem: A Supply-Side Record of Agentic Automation (8 October 2026, 00:00 UTC)
not affiliated with or endorsed by Cohere