WebMCP

This site implements the Web Model Context API, so an AI agent in the browser can call tools instead of scraping the rendered page. Everything is generated at build time and served as static files. There is no server and no key.

The tools

Registered on every page of the site. Individual pages can add their own on top, as both datasets below do.

Tool	What it does
`search_site`	Keyword search across posts, projects, art and pages
`list_posts`	Posts newest first, with paging
`get_post`	Full markdown of one post by slug
`list_projects`	Playground projects, optionally filtered by tag
`list_art`	Art pieces
`get_current_page`	Title, headings, links and text of the page on screen
`get_contact_options`	Booking link, LinkedIn, X, GitHub, shop, llms.txt

Using it without WebMCP

The tools need a browser that exposes document.modelContext. If yours does not, the same data is plain JSON over HTTP and needs no client at all:

/webmcp/insolvencies.json the insolvency ledger, its aggregates and the national trend
/webmcp/robotics.json the humanoid supply chain watchlist: every company, its prices and returns, the correlations, the claim audit and the log
/webmcp/manifest.json every post, project, art piece and page, plus a datasets array pointing at the queryable ones
/webmcp/posts/<slug>.json the full markdown of one post
/llms.txt the plain-text site description

Everything is read-only. All tools are annotated readOnlyHint, and the ones returning ledger prose or page text carry untrustedContentHint, since that text comes from news sources rather than from me.

Datasets

Two of them. Both are scoped to their own page: WebMCP is per-document, so an agent reading a blog post should not be offered nine tools about German insolvencies and humanoid robotics. They register on the pages below and nowhere else. An agent elsewhere finds both in the datasets array of the manifest, which carries the data URL, and can either fetch that file directly or navigate to the page and pick up the tools.

The insolvency ledger

Germany bites the dust is a running ledger of 135 named German company insolvencies since February 2025, 53,970 jobs between them, each with a published headcount and a linked source. Four tools query it, and the whole dataset is one file:

/webmcp/insolvencies.json

Tool	What it does	Arguments
`insolvency_summary`	Totals and breakdowns: cases, jobs, how many have stopped trading, and the split by sector, state and month. Start here.	None
`list_insolvencies`	The cases themselves, newest first, paged. Filter by sector, state, date range, minimum headcount, or whether the business has stopped.	sector, state, since, until, minEmployees, ceasedTrading, sortBy, limit, offset
`get_insolvency`	The full record for one company, including the narrative detail, what the headcount counts, and the source link.	company
`insolvency_trend`	The national monthly series from the IWH-Insolvenztrend: filings and jobs affected across all German partnerships and corporations.	None

What the data does and does not claim

A case is listed only if a named source published a specific employee headcount, a real town and federal state, and a defensible date. Cases without a published headcount are left off rather than estimated, which is why some genuinely large failures are missing.
It is a sample, not a register. Germany logged roughly 12,900 corporate insolvencies in the first half of 2026 alone. A company's absence here is not evidence it is solvent.
Filing is not closing. Most listed companies are still trading; ceasedTrading marks the ones that have actually stopped.
Plant closures, restructurings and reform programmes are excluded, however many jobs they cost, because they are not insolvencies.
The national trend comes from the IWH-Insolvenztrend, which counts a far wider universe than this ledger. Do not sum the ledger to estimate a national total. IWH publishes each month in the second week of the following month, so the latest month available is August 2026.

The humanoid supply chain watchlist

Humanoid supply chain prices and fact-checks 17 companies across three tiers: the component makers who sell to every assembler, the rare-earth and critical-minerals names they get confused with, and the magnet layer in between. 13 are listed; the rest are private and are carried because they move the thesis, not because you can buy them. Five tools query it, and the whole dataset is one file:

/webmcp/robotics.json

Tool	What it does	Arguments
`robotics_summary`	The three tiers, what each one is, how many names are listed against how many are private, and every correlation to REMX and to the robotics basket. Start here.	None
`list_robotics_companies`	The priced rows: last close, one day, one month, year to date, one year and how far below the 52-week high. Filter by tier or to listed names, sort by any return column.	tier, listedOnly, sortBy
`get_robotics_company`	One company in full: price, returns, correlations, the verified figures, the one thing to watch, and on request the written analysis.	company, detail
`robotics_briefing`	The dated running log, newest first, plus the catalysts worth watching and which way each one cuts.	entries
`robotics_claim_audit`	The fact-check: claims made about these companies in viral posts and the press, each with a verdict and what the filings say.	query

What the data does and does not claim

It is a research snapshot, not investment advice, and it makes no recommendations. Every figure is as of 26 Sep 2026, which the payload carries as compiled so an agent can say how old it is before quoting it.
Prices are delayed closes, not live quotes: each is its venue's last completed session in the listing's own local currency, so the percentage moves carry no currency translation and the levels are not comparable across venues. Tokyo is usually a session ahead of Xetra and New York.
A null ticker means the company is not directly investable. access says how you would own it, if at all, and any price on such a row belongs to the parent named in pricedVia rather than to the company itself.
Correlations are weekly returns over the last 104 completed ISO weeks, bucketed by exchange-local date so Tokyo, Xetra and New York weeks line up. The robotics basket sits at about 0.4 to REMX: real co-movement, not a proxy, and not causation.
The claim audit is the part worth calling: 27 claims made about these companies in viral posts and in the press, checked against filings. Several are wrong, including a ticker that buys a different company.