MCP connector to the data

Everything the site shows, you can query yourself. An MCP server (Model Context Protocol) gives Claude, Codex, Kimi, GLM and your own scripts the same numbers as tools: the time series behind the index, every occupation with its tasks and Claude use, and every field of study. All of it is aggregate data that is already public. No login.

Connect

The endpoint is https://kiindeksen-mcp.fly.dev/mcp (Streamable HTTP, JSON-RPC 2.0).

Claude Code. One command:

claude mcp add --transport http kiindeksen https://kiindeksen-mcp.fly.dev/mcp

claude.ai. Settings, Connectors, Add custom connector. Paste the address above. Then ask Claude for, say, "the headline number with uncertainty for those under 31" or "compare occupations 4110 and 3312".

Your own script. client_example.py shows the whole exchange with the server and needs only Python, no packages:

python3 client_example.py

Other clients and models

The connector is open MCP over HTTP and is not tied to Claude. Any client that takes a URL can use it, and the URL is the same everywhere: https://kiindeksen-mcp.fly.dev/mcp.

From code, with any provider

The OpenAI and Anthropic model APIs call the server directly, without an MCP client of your own. With OpenAI (Responses API) the tool is {"type": "mcp", "server_label": "kiindeksen", "server_url": "https://kiindeksen-mcp.fly.dev/mcp", "require_approval": "never"}. With Anthropic (Messages API, beta header mcp-client-2025-11-20) the server is {"type": "url", "url": "https://kiindeksen-mcp.fly.dev/mcp", "name": "kiindeksen"} in mcp_servers, and the tool is {"type": "mcp_toolset", "mcp_server_name": "kiindeksen"}.

Any other model with function calling, such as Kimi, GLM, Gemini or Mistral, can use the same tools as plain HTTP calls. Every tool exists as GET /tools/<name> with the arguments in the query string (lists comma-separated) or as POST with a JSON body. openapi.json describes them, so GPTs, Dify, Coze and similar platforms can import them. Try:

curl "https://kiindeksen-mcp.fly.dev/tools/get_headline?young=true&reference=claudecode"
curl "https://kiindeksen-mcp.fly.dev/tools/compare_occupations?codes=4110,3312&index_base=2025-02"

The tools

Two resources (kiindeksen://om, kiindeksen://konvensjoner) and two prompts (analyser_yrkespar, maanedsrapport) come with it.

Conventions

• The series are indices with 100 in November 2022, the month ChatGPT arrived. Use index_base to move the base, for example 2025-02 for Claude Code.
• sa is seasonally adjusted, percap is per resident in the age group.
• The index is growth in the most exposed fifth against growth in the least exposed, in percentage points. get_headline gives the site's number with uncertainty. get_growth lets you compute your own references and smoothing.
• Occupations are four-digit STYRK-08. Only occupations with at least 30 employees every month have a series. Automation is Anthropic's rule: directive plus feedback loop.
• Fields of study are three-digit NUS groups, level plus field.
• State the release, specification and base month when you report numbers.

Example: same exposure, different automation

The question: two occupations have almost the same AI exposure. Is AI used the same way in them, and does employment move the same way? This is what client_example.py does, and what Claude does if you ask for the prompt analyser_yrkespar with codes 4110 and 3312.

1. get_release and get_headline. Release 2026-10 covers January 2021 to July 2026. The index for ages 21–60 with ChatGPT as reference is +1.5 percentage points, standard error 2.7. For ages 21–30 since Claude Code it is −2.9, standard error 2.0.
2. get_growth on by_exposure with reference 2025-01 and three-month smoothing. From January 2025 to July 2026, seasonally adjusted: least exposed fifth −0.3 percent, most exposed −0.3 percent. The gap is −0.1 percentage points. The same call on age_by_exposure for ages 21–30: −1.8 in the least exposed, −4.6 in the most exposed.
3. compare_occupations with codes 4110 and 3312, employment, seasonally adjusted, base month 2025-02:

   Occupation	Eloundou	Mouchel	Automation (Feb 2026)	Employment July 2026
   4110 General office clerks	0.59 (Q5)	0.59 (Q5)	67 %	95.7
   3312 Credit and loans officers	0.61 (Q5)	0.62 (Q5)	30 %	96.5

4. get_occupation for each occupation, platform claude_ai, gives the tasks. Office clerks: troubleshooting office equipment (1.26 percent of all Claude use, automation 60 percent), forms and expense reports (66 percent) and office machines (51 percent). Credit officers: recommending financial strategy (24 percent), explaining credit assessment and regulation (22 percent) and explaining loan types (27 percent).
5. The interpretation is yours: the same exposure score, but in one occupation the work is ordered finished, in the other it is discussed. That is what the exposure measure does not see and the usage data shows.

The same exercise for a field of study: search_education with "økonomi" at bachelor level gives group 641 Business and administration. It has 101 315 employed, exposure 0.55, half in the most exposed fifth and automation 43 percent. get_education_group gives the ten most common tasks in the jobs and the most common jobs, accountants first.

All numbers above were read from the server on 15 October 2026 with release 2026-10. New releases give new numbers. Run the script again.