Public MCP directories contain nearly 700,000 tool definitions.

But their overlap with workplace tasks is much narrower.

We mapped public MCP tool definitions to workplace tasks to see where developers are concentrating their efforts.

How MCP works

A customer service representative receives a complaint about a missing order. The representative confirms that it was delayed in customs and is expected by the end of the week. After resolving the delayed order, they ask an AI assistant to record the interaction and resolution in their support software.

How a requested software action moves through MCPA six-step vertical flow. A person requests an action. An AI application identifies the requested action, selects the appropriate tool definition from available actions, and organizes the relevant details into labeled pieces. The selected action feeds into the assembled request. The MCP server validates and passes that request to support software, which records the resolved interaction. The result returns through the server to the application and representative.CUSTOMER SERVICE REPSends a requestDescribes the intended outcomeAI APPLICATIONIdentifies the actionIdentifies the requested software actionMCP SERVERAVAILABLE ACTIONSRecord interactionSearchUpdateASSEMBLED REQUESTActionRecord interactionIssueMissing orderResolutionDelivery expected this weekValidate requestPass to softwareSUPPORT SOFTWARERecords the interactionPerforms the requested actionSTRUCTUREDRESULTRecordconfirmed

  1. Step 1: The representative sends a request

    The representative asks the AI application to record a resolved customer interaction.

  2. Step 2: The AI application identifies the requested action

    The AI application identifies that the representative wants to record the resolved interaction in support software.

  3. Step 3: The AI application selects an MCP tool definition

    From the available actions exposed by the MCP server, the application selects “Record interaction.”

  4. Step 4: The AI application organizes the request

    The application organizes the action, issue, and resolution into predictable labeled pieces.

  5. Step 5: The MCP server passes the request to support software

    The MCP server receives and validates the request, then passes it to the support software, which records the interaction.

  6. Step 6: The AI application presents the result

    The result returns through the MCP server. The AI application presents it to the representative.

Active step: The representative sends a request

Six steps show how an AI application selects an available action, organizes the relevant details into a predictable request, sends it through an MCP server to support software, and presents the returned result to the representative.

MCP’s significance is that it replaces a separate custom integration for every pairing of an AI application and a software system. Developers can publish an action once through an MCP server, and many compatible applications can discover and use it through the same protocol.

That shared standard makes software capabilities easier to build, distribute and reuse. The directories we analyzed contained 123,069 server listings and nearly 700,000 tool definitions, spanning actions from searching databases to editing files and generating images. At that scale, a new question emerges: what kinds of work are developers building for?

One software action can overlap with one piece of a job

The customer-service example shows one narrow action: recording a customer interaction and its resolution in support software. O*NET, a database sponsored by the U.S. Department of Labor, describes occupations through their individual tasks. For Customer Service Representatives, one of those tasks is keeping records of customer interactions.

The MCP tool definition and O*NET workplace task describe similar work from opposite sides. One publishes a software action. The other records what a person does as part of their job.

One accepted description-level match between an MCP tool definition and an O*NET workplace task

O*NET WORKPLACE TASK

“Keep records of customer interactions or transactions, recording details of inquiries, complaints, or comments, as well as actions taken.”

Customer Service Representatives

MCP TOOL DEFINITION

mind_crm

Records calls, emails, meetings, and notes in a contact history.

We counted an accepted tool-task match only when the tool description directly covered the action described by the O*NET workplace task. Of nearly 700,000 tool definitions, 2.6% met that strict definition. Together, the accepted matches covered 1,380 of 9,423 target occupation-task pairs across 491 of the 895 occupations we studied.

Accepted matches were highly concentrated. The ten most-targeted tasks accounted for 34% of all accepted task-tool pairings among occupations included in the chart.

The chart shows how many MCP tool definitions matched each workplace task. Each circle represents one task; larger circles farther left indicate more matched tool definitions. Select a circle to see the task and its exact match count on the right, where each blue dot represents one matched tool definition.

Each blue circle represents one workplace task. Move the pointer to inspect the nearest task. Click or tap to lock. Use arrow keys, Home, End, Enter, Space, and Escape to move through tasks.

MORE MATCHESFEWER MATCHESMATCHED TOOL DEFINITIONS · LOGARITHMIC SCALE
Use computer software to generate new images.

Use computer software to generate new images. Occupation: Graphic Designers. 990 matched tool definitions. Rank 1 of 1,361 tasks.

The result is a steep head and a long tail. The leading task, using computer software to generate new images, matched 990 tool definitions. The median task matched only 2, and nine out of ten tasks matched 18 or fewer.

Public MCP development is therefore clustering around a small number of workplace actions rather than spreading evenly across tasks. The same imbalance appears at the occupation level: accepted matches reached 473 occupations, but tool definitions accumulated heavily in a much smaller group.

Most workers are in occupations with shallow MCP tool coverage

Each circle is one occupation with at least one accepted tool-task match. Occupations farther right have a larger share of all O*NET tasks matched. Occupations higher up have more total matched MCP tool definitions. The circles then change size to show U.S. employment.

Most workers are in occupations with shallow MCP tool coverage

Each circle is one occupation with at least one accepted tool-task match. Occupations farther right have a larger share of all O*NET tasks matched. Occupations higher up have more total matched MCP tool definitions. The same circles then change size to show U.S. employment.

Matched tool definitions and U.S. employment follow different patterns. The contrast between Customer Service Representatives and Graphic Designers makes that difference concrete.

Customer Service Representatives

Matched tool definitions
3
U.S. employment
2,595,750
Share of all O*NET tasks matched
15.4%

Graphic Designers

Matched tool definitions
1,061
U.S. employment
197,830
Share of all O*NET tasks matched
57.9%

Customer service employs about 13 times as many people, while graphic design has more than 354 times as many matched tool definitions.

Task coverage and matched MCP tool definitions across occupationsHorizontal position shows task coverage, vertical position shows matched tool definitions, and circle area can show May 2025 U.S. employment.
10 largest directly estimated workforces10 occupations with the most matched tool definitionsOther occupations
The chart plots 473 occupations with at least three software-performable O*NET tasks and at least one accepted match. U.S. employment uses May 2025 direct BLS estimates.

The contrast between Customer Service Representatives and Graphic Designers helps explain this pattern. For Customer Service Representatives, the clearest connection was recordkeeping in support software, a task already carried out in structured software. For Graphic Designers, matched tool definitions concentrate around image-generation platforms, where the work is digital from start to finish.

These examples are consistent with the ecosystem’s origins in software. Tasks involving code, databases, APIs and digital files are already legible to software. That legibility likely contributes to technical occupations standing out in the data.

The public MCP ecosystem is broad but uneven. Developers have concentrated tool supply in tasks already expressed through code, databases, APIs and digital files. Many of the country’s largest workforces sit elsewhere. Workplace adoption is a separate question. Where organizations adopt these tools will shape which parts of jobs change and what expertise remains important.

Explore how AI could reshape expertise at work →

Cohere Labs