Using the Data From Your AI Assistant

What this is

This site publishes its data over the Model Context Protocol (MCP) — an open standard that lets an AI assistant such as Claude look data up from an external source while you chat with it. You add one URL to your assistant; after that, questions about Japanese prices are answered from this site's numbers rather than from the assistant's memory.

The assistant does the reading and writing. The server only answers data lookups, and it can only see what this site already publishes: the same figures, from the same accepted release, computed by the same code as the public API. Nothing you ask is used to change anything here — every lookup is read-only.

Getting connected

The connect URL is:

claude.ai and Claude Desktop

1. Open Settings → Connectors (a paid Claude plan is required for custom connectors).
2. Choose Add custom connector, name it Japan Data Observatory, paste the URL, and add it. No authentication is asked for — there is none.
3. Start a new chat and confirm the connector is enabled in the tools menu under the message box. Connectors added mid-chat are often not picked up until a new conversation.

Claude Code (terminal)

Run it once; every new session then has the tools. claude mcp list should show the server as connected.

Everything else

Any client that supports remote MCP servers (Streamable HTTP transport) takes the same URL. The server is stateless, needs no key, and identifies itself as Japan Data Observatory.

The lookups, one by one

You never call these yourself — the assistant chooses them from your plain-English question. Knowing what exists tells you what you can ask.

Consumer prices (always available)

• List datasets — the published tables and what each covers. The assistant uses it to orient itself.
• Search series — finds any of the ~820 CPI series by English or Japanese name, or by series code. Returns each match with its basket weight, latest index value, and latest year-over-year and month-over-month rates.
  Ask: “Is there a series for mobile phone charges?”
• Get series history — the history of up to six series at once, as the published index level (2020 = 100) or as a year-over-year, month-over-month, or 3-month annualized rate, over any date window back to January 1970.
  Ask: “Rice prices year-over-year since 2022.”
• Inflation overview — headline, core (less fresh food), and core-core (less fresh food & energy) at the latest month, plus the ten major expenditure groups with their weights.
  Ask: “Where is Japanese inflation right now?”
• Contributions to headline — the percentage-point decomposition of headline year-over-year inflation by expenditure group, with the rounding residual disclosed separately.
  Ask: “How much of June’s inflation was food?”
• Inflation breadth — across the 582 individually priced items: the share rising at least a threshold, the share rising at all, and the share falling.
  Ask: “What share of items is rising more than 2%?”

Cross-shareholdings (where published)

A second dataset — Japanese policy shareholdings (政策保有株式) extracted from annual securities reports filed on EDINET — is being built filing by filing. Its lookups appear automatically on servers where the dataset is published; if your assistant doesn't list them, the server you're connected to doesn't carry the data yet.

• Cross-shareholding coverage — how many filers are extracted so far, named holdings, total book value, and positions reduced versus increased. The dataset is not yet the whole market, so any aggregate should be quoted with this denominator.
  Ask: “How much of the cross-shareholding data is covered so far?”
• Search companies — find a company by Japanese name or securities code, with how many holdings it files and how many filers hold it.
• Company holdings — one company in both directions: every named holding it discloses (shares, yen book value, prior-year figures, the stated purpose as filed, reciprocity) and every extracted filer that holds it. The reverse view exists nowhere in any single filing — it only exists because every filer's table is in one database.
  Ask: “Who holds Mitsubishi UFJ shares as policy holdings, and why do they say they hold them?”
• Unwind ranking — extracted filers ranked by total named policy-holding value, with prior-year value and reduce/increase counts.
  Ask: “Which filers hold the most, and who is actually reducing?”

How to read what comes back

Every lookup result carries the same disclosure fields. If you ever need to check the assistant's statement, these are what to look at:

• trust — official means the figure is exactly as published by the source (index levels; filed share counts and book values). calculated means this platform computed it from official inputs.
• calc — for every calculated figure, the exact formula used. A rate computed from published (rounded) indices can differ from the source's own published rate by up to ±0.1 pp; the Methodology page explains why.
• cite — a permanent URL on this site showing the same view. Click it to see the numbers the assistant saw. This is the link to quote in a note or a post.
• as_of_release / latest_period — which data release answered the lookup. The data is monthly; a figure is always “as of” a named month.
• Missing values come back as missing, never as zero. If the assistant tells you a number is unavailable, that is the data speaking, not a failure.
• Coverage (cross-shareholdings) — extraction status and filer counts ride along with equity responses. “¥21tn across 125 filers extracted so far” is honest; “Japan's cross-shareholdings total ¥21tn” is not, and the data says so.

Ten questions to try

1. What are Japan's headline, core, and core-core inflation rates right now?
2. Break the current headline rate into percentage-point contributions by group.
3. Is inflation accelerating? Compare year-over-year with the 3-month annualized rate.
4. How has the price of rice moved over the last three years?
5. Compare electricity and city gas since 2022 on one timeline.
6. What share of individually priced items is rising more than 2%? How does that compare with two years ago?
7. What happened to mobile phone charges in 2021?
8. With the cross-shareholding data: what's the coverage so far?
9. Who holds company 8306, and what reasons do the filers state?
10. Rank the extracted filers by policy-holding value and tell me who reduced positions year-on-year.

Limits, honestly stated

• Consumer prices and cross-shareholdings only. No Bank of Japan balance-sheet data, wages, GDP, or producer prices yet; no Tokyo advance, regional, or seasonally adjusted CPI series; no yen price levels (an index says rice rose 30%, not that it costs ¥3,000). The coverage grows dataset by dataset.
• Latest published data only. “What did April look like before the revision?” is not answerable yet — point-in-time history is stored and will be exposed in a future release.
• History windows are trimmed. Long histories return the most recent points within a size budget; the assistant can page through older windows by asking with explicit dates. Up to six series per history lookup.
• Rate limit. Roughly 120 lookups per minute per connection — far more than a conversation uses. Bulk work belongs on the public API.
• The assistant's prose is its own. This server hands over numbers with labels and links. Interpretation, forecasts, and causal stories are the product of whichever AI you connect — this site publishes none of those. Check anything you plan to quote against its cite link.

When something looks wrong

• The connector won't add — custom connectors need a paid Claude plan; free accounts can't add them.
• The tools don't appear — start a new chat (web/desktop) or a new session (Claude Code); check the connector is enabled in the chat's tools menu.
• “Rate limit exceeded” — wait a minute. If a single question routinely trips it, the assistant is looping; rephrase more narrowly.
• The assistant's rate differs slightly from an official publication — expected within ±0.1 pp; see Methodology → Calculations. Index levels should never differ; if one does, that is a bug — please report it.
• Cross-shareholding lookups are missing or answer “not published yet” — the server you're connected to doesn't carry that dataset yet.
• A number with no cite link, or a claim you can't reproduce — treat it as the assistant's own words, not this site's data.

For data teams

The endpoint speaks JSON-RPC 2.0 over a single stateless POST — no session, no authentication, gzip supported. A minimal round-trip:

curl -X POST https://web-production-c9178.up.railway.app/mcp -H "Content-Type: application/json" -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Tool results return the same JSON the site's public API serves, plus the disclosure fields above. For scheduled or bulk retrieval, use the REST API directly — it is the same data with proper HTTP caching.