Overview
How the website works, the webapps it contains, the flywheel of flywheels, WebMCP and the API, and the Hive; then where to start.
Cedwyn is one website at cedwyn.bucklerfamily.estate made of many separate web apps, called sections. Each section lives on its own branch of one Git repository and is published under its own path, for example /wayback/, /maps/ or /research/. A shared header lets a visitor move from one section to the next.
Under the sections sits one core catalogue database: a single SQLite file built on every deploy from everything the sections publish. It holds every file once (by its sha256), where each file appears, a full-text index and a graph of links. Any page or tool can read it in the browser, with no server.
How the whole site works: one flywheel of flywheels
BucklerFamily.Estate is the brain of a long piece of research: an LLM wiki and evidence catalogue where every record is collected, transcribed, described and interpreted, and where the research is reported as it goes, in the public interest. Keeping it running is one giant snowballing flywheel, the Cedwyn Snowball. Each turn of the wheel makes the record more accurate, and a more accurate record makes every output better.
- Ask. Research questions and records requests go out: freedom-of-information and subject-access requests, archive catalogues, books and online sources.
- Receive. Replies, disclosures and original documents come back. Refusals and silences are recorded too, with their dates.
- Clarify. Every answer raises more questions. Clarification is sought, and the best contrary reading is tested.
- Organise and officiate. Answers organise the knowledge base: each record gets one catalogue entry (a neutral description, a verbatim transcript, a label, its parcel and its date), and findings become official records with their sources.
- Reconstruct. The story, the estate, the family and the public history and heritage are rebuilt into one central, chronological, accurate record of the Cedwyn estate, the former manorial house and the Williams families.
- Present. That record feeds a better museum in Argentina, a guided tour in Wales, the online dimension (maps, games and other webapps that help people make sense of a huge and complex story), documentary scenes made in real time, and e-books.
- Re-release. Every output can be re-released under a new semantic version as the quality of the information improves; earlier versions stay listed with a changelog.
- Learn and return. Readers, replies and new finds raise the next questions, and the wheel turns again.
The webapps
The site is about fifty webapps on their own paths, each built from its own branch: the evidence layer (research catalogue, evidence catalogue, records, wayback copies, watchtower, title chain, tithe map, names), the story layer (story, e-books, documentary, blog), the online dimension (maps, terrain, games) and the programme layer (projects, calendar, directory, shop), with Welsh, Argentine and online scopes. The shared header numbers each one as a slot of 59, one for each plot on the 1824 survey plan. Every webapp reads from the central record rather than keeping its own copy of the facts. The full list is under Sections.
Flywheels inside the flywheel
Smaller flywheels turn each arc on a schedule: request drafting, correspondence and contacts (ask); email ingest, first-seen and page recapture (receive); the reopen and intelligence loops (clarify); task-register burndown and repository sync (organise); the research, parcel and webapp-migration loops (reconstruct); documentary scenes, the programme office, social and growth loops (present); the version standard and releases (re-release); and the Agent Library and skills loops (learn). Underneath runs the upkeep: the daily site check, the problems queue, builds, Cloudflare, GitHub and the task register.
WebMCP, the API and the Hive
Software agents use the same site without scraping it. Every page loads read-only WebMCP tools for searching the catalogue and reading records, with one submission route whose items always wait for review. Two MCP servers, a static data API and an in-browser database complete the API. Select agents collaborate in the Hive: they claim one job at a time, make one checked change, sign it, and hand on; draft letters need independent reviews of the exact same version, and only the owner sends them.
Keeping it running
The maintenance and upkeep handbook (version 0.2.0) covers the repository and branches, Cloudflare, GitHub, the task register, the Hive, the version standard, publishing and the rules for public text. Agents read it at /academy/handbook/; the programme itself is described on Cedwyn Snowball.
Where to start
- Architecture: the core database, section data and how records are shared without copies.
- Hosting: the GitHub repository, the build workflow and Cloudflare Pages.
- Front ends: static, serverless pages; layout in the section, data from the database.
- Sections: every section, its branch, its data and its status.
- Add a section: a new web app on a new branch, step by step.
- Data export: the whole database as CSV, JSON and SQLite.
- API and WebMCP: endpoints, the in-browser SQL and GraphQL interface, and tools for browser agents.
- Cedwyn Snowball: the programme-wide flywheel for estate reconstruction, records recovery, documentary production, agent coordination, business development and future heritage.
- Narrator voice: cloning method and profile for the narrator voice in the video intros.
- Status: what is built today and what is not yet.
Design principles
- One record, many views. A file is stored once and referenced wherever it appears.
- Sections are independent. A section builds, deploys and rolls back on its own branch. Changing one section does not change another.
- Static first. Everything a visitor loads is a plain file. No application server, no per-request cost.
- Portable data. The whole database can be downloaded and opened with ordinary tools, separately from any front end.
- Pages do not explain themselves. How the site works is documented here, in the library; human pages carry the record only. Notes for software agents travel in page metadata,
/agents.json,/llms.txtand/academy/.