English summary
Daily supply-and-demand data for service marketplaces.
MarketKernel observes service marketplaces — Coconala, Fiverr, Upwork and others — once a day, and normalises demand, supply and price into one comparable scale across Japan, the US and Europe. We publish observed facts and the method behind them. We do not publish rankings, revenue forecasts or recommendations.
Status: observing since 2026-08-25. Provisional indices at day 30, change decomposition at day 60, the first index at day 90. Every figure currently shown on this site is a formatted sample, not a measurement — except on the research page, where each claim carries its source. This page is a summary; the full site is in Japanese.
The problem
Marketplace numbers don't add up — literally.
A listing count on Coconala, a Gig count on Fiverr and a job-posting count on Upwork are different units. Summing them produces a number that means nothing. And a snapshot taken today cannot tell you what changed over the last 90 days: only someone who started measuring 90 days ago can answer that. MarketKernel exists to do the boring part — normalise the units, and keep measuring every day.
What you get
Seven numbers per market, never collapsed into one score.
A single composite score hides scale: 100 ÷ 10 and 10,000 ÷ 1,000 give the same ratio. We keep the relative index and the absolute volume side by side, and we never hide low-confidence rows.
| Field | Meaning | Unit |
|---|---|---|
demand_index | Buyer-side activity, from sales velocity and review growth | 0–100 |
supply_index | Seller-side volume, from listings, bids and proposals | 0–100 |
gap_index | demand − supply. Positive means buyer-side, negative means seller-side. Not a judgement of whether the market is good | ±pt |
momentum_90d | Change in the gap over the last 90 days | ±pt |
price_median, price_p25/p75 | Median price and quartiles. Stored in USD, displayed in JPY with the FX rate attached | currency |
volume_absolute | Observations per month. Not transaction value. Never omitted | obs/mo |
confidence, volume_confidence | How solid the figure is, from signal quality, source multiplicity, sample size, observation length and freshness. Rows below 60 are shown, not hidden | 0–100 |
Demand and supply changes are reported separately, so a moving gap can be attributed to buyers arriving or sellers leaving — not just to the gap itself. Employment demand (job postings) is kept in its own layer, never mixed into outsourcing demand.
How it is built
Collect, align, hand over.
CollectPublic pages only, once a day, at low frequency. Listing counts, request counts, review growth and prices are recorded as raw objects and hashed.
AlignSurface expressions in any language — "ロゴ制作", "logo design", "diseño de logo" — are mapped onto one market definition. The mapping table is versioned and frozen; no LLM classifies anything at runtime.
Hand overPlatform-internal relative values are calibrated against anchor categories to recover absolute volume. The calibration version and its estimation error ship with every value.
Crawling policy
If we can't take it without evasion, we don't take it.
- Public, non-logged-in pages only. One pass per day, sequential, low rate
robots.txtfetched and honoured on every run, and the result recorded- Identifying User-Agent with the operator name and a contact URL. Never anonymised
- On 429 or 503 we stop for the day and do not retry
- No CAPTCHA solving, no browser-detection evasion. Sites that require evasion are dropped from self-collection even when the data would be valuable
- No seller names, profiles, avatars, transaction records or messages
If you operate a marketplace and want to be excluded, a robots.txt rule is enough — or write to us and we will remove you and record the exclusion publicly.
Every source is published
All 90 sources are listed with their region, signal type, collection method, cadence and status — live, connecting, preparing, reading-only, deferred or excluded — including the ones we decided not to use and why. Seven are live today; the rest are still being connected or checked. Sites that returned 403 to our identifying User-Agent are listed as such.
Auditability
Every value traces back through the index formula, the calibration coefficient, the ontology mapping, the observation, and finally the SHA-256 of the raw page. The design chains each daily manifest to the previous day's head and anchors it externally, so that "we did not rewrite history" becomes verifiable without trusting us. The genesis record exists; chain heads and external anchors will be recorded from the first daily run onward.
API and MCP
Built to be called by agents, not only read by people.
The same core serves a REST API (/v1) and an MCP server (Streamable HTTP, stateless). Six read-only tools — search_markets, get_market, get_market_trend, compare_markets, get_market_sources, trace_market_value — plus resources for the methodology, the ontology, the source registry and the metric definitions.
compare_markets(
markets=["AI_AGENT_DEVELOPMENT"],
regions=["JP","US","EU"],
metrics=["demand","supply","gap","price_median","confidence"]
)
→ by_region.JP demand 84 / supply 51 / gap +33
by_region.US demand 88 / supply 62 / gap +26
cross_market_gap demand_level_diff −4
note "Facts placed side by side. No interpretation."Interpretation stays out of the tools
The server instructions and the marketkernel://policy resource state plainly that no tool returns entry recommendations, revenue forecasts or overall rankings. Search results are always sorted by market name. volume_absolute and confidence are never omitted, and trace_market_value returns observed lineage only.
Response schemas are frozen at v0.1: fields may be added, never renamed or removed without a version bump.
Endpoints are not open yet. The specification is published so you can evaluate it now; access opens alongside the first real dataset. If you want to connect, tell us which markets you need.
Where we are
Nothing here is measured yet — and that is the point.
Observation began on 2026-08-25. Historical series cannot be created retroactively, so the only way to have 90 days of history is to have started 90 days ago. Until the first dataset is published, every number on this site carries sample: true, and the API returns the same flag.
Coverage today
32 market definitions across Japan, the US and Europe. Categories with no counterpart in a region are returned as "no counterpart" rather than force-mapped.
Licensing
Time-limited update rights, tiered by what you may do with the data — internal analysis, client deliverables, public attribution, or API and agent access. We do not sell exclusive access to the data itself.
Not for individuals
We do not sell directly to individuals starting a business. Neutral data is only useful to someone who can interpret it, and we do not supply the interpretation.
Tell us which markets you need observed.
We are currently taking five partners who specify the categories we observe. Observation of your categories starts the next business day; the first index lands 90 days later. Press enquiries, data partnerships and marketplace operators asking to be excluded are all welcome at the same address.