MCP 服务器

Wheel of Heaven MCP Server

A read-only MCP server that lets AI agents search and retrieve the Wheel of Heaven corpus, including source-grounded facts, interpretations, and comparative traditions, all with full epistemic metadata.

README

mcp.wheelofheaven.world

Read-only MCP server for the Wheel of Heaven corpus. It wraps the static JSON API at api.wheelofheaven.world so AI agents can search and retrieve the corpus with full epistemic metadata (claim types, editorial passes, citations).

Design rationale and tool-surface spec: .claude/plans/strategy-mcp-server.md in the org .claude repo.

Endpoints

/mcp — Streamable HTTP (canonical)
/sse — legacy SSE transport
/ — server info (JSON)

Tools

Fact layer — source-grounded data, no interpretation applied:

Tool	Purpose
`search_corpus`	Fuzzy full-text search over wiki, timeline, library books, articles, and news (~290 documents; token-AND with OR fallback)
`get_entry`	One wiki/timeline/articles/news entry as markdown with `claim_type`, `editorial_pass`, typed related-content edges, and canonical citation URLs; 9-language support via `lang`
`get_passage`	A chapter of a digitized primary source text as numbered paragraphs with stable reference IDs (e.g. `GEN-1:26`); `ref` filters to one verse
`get_source`	Bibliography lookup by stable ID or search — every record carries the source-program classification (family, stance, relation), critical sources included
`compare_traditions`	Rows from the comparative datasets: flood-myths, divine-council-index, theomachy-crossrefs, world-ages, prophets-and-religions
`query_graph`	Typed 1-hop ego network of any corpus node (`see_also`, `in_body`, `cites_source`, `same_tradition`, `same_age`, `comparison_of`)

Interpretation layer — explicitly the project's own reading, labeled as such:

Tool	Purpose
`get_interpretation`	The Wheel of Heaven framework's reading of a topic (curated context documents, or the closest wiki entry), always carrying an interpretation-layer notice
`get_method`	The editorial methodology: claim-type taxonomy, six-source rule, editorial passes — how to weigh everything else

Meta:

Tool	Purpose
`get_glossary_term`	Core terminology with renderings in all 9 site languages (exact-id match first)

Resources: the six /v1/context/* narrative documents (woh://context/{overview,hypothesis,terminology,timeline,sources,method}) plus woh://llms.txt.

Development

npm install
npm run dev        # wrangler dev on :8787
npm run typecheck
npm run deploy     # wrangler deploy (needs Cloudflare auth)

The server is stateless over the upstream API: every tool call fetches api.wheelofheaven.world/v1/…, with the ~600 KB search corpus cached in-memory per isolate for one hour (matching the API's edge-cache TTL).

Architecture

Cloudflare Workers + agents SDK (McpAgent Durable Object)
src/server.ts — transport-agnostic tool registration (reused by the future npm stdio build)
src/api.ts — /v1/ client + search-corpus cache
src/html.ts — deterministic body_html → markdown conversion