MCP Server
The Wheel of Heaven MCP server — connect AI agents to the corpus via Model Context Protocol: 9 read-only tools with claim-type labels and an explicit fact/interpretation split.
The Wheel of Heaven MCP server exposes the corpus to AI agents over the Model Context Protocol. It is a thin, read-only wrapper around the static JSON API: every tool call fetches api.wheelofheaven.world/v1/… and reshapes the response for agent consumption, with epistemic metadata (claim types, editorial passes, citation URLs) attached to everything.
| Endpoint | https://mcp.wheelofheaven.world/mcp (Streamable HTTP) |
| Legacy transport | https://mcp.wheelofheaven.world/sse (SSE) |
| Auth | None — the corpus is CC0, the server is read-only |
| Source | github.com/wheelofheaven/mcp.wheelofheaven.world |
| License | CC0-1.0, like the corpus itself |
Quickstart#
Claude Code
claude mcp add --transport http wheel-of-heaven https://mcp.wheelofheaven.world/mcpClaude.ai / Claude Desktop — add a custom connector with URL https://mcp.wheelofheaven.world/mcp (no authentication).
Cursor / other Streamable-HTTP clients — point the client at the same URL; the server is stateless and requires no headers beyond the MCP defaults.
Local stdio (npm) — for clients that prefer a local process:
npx -y @wheelofheaven/mcpThe server is also listed in the official MCP registry as world.wheelofheaven.mcp/corpus, so registry-aware clients can discover it without any of the above.
Design: two layers, declared stance#
The tool surface enforces the corpus’s editorial split (see Editorial method): fact-layer tools return source-grounded data with no interpretation applied; interpretation-layer tools return the Wheel of Heaven framework’s own reading and say so in every response. An agent can use the comparative datasets without ever invoking the interpretation — and when it does invoke it, the framing is self-declared rather than smuggled.
Every response carries claim_type labels (direct / framework / inferred / speculative), the canonical www.wheelofheaven.world URL for citation, and the /v1/ JSON URL.
Tools#
Fact layer#
| Tool | Returns |
|---|---|
search_corpus | Ranked fuzzy search over wiki, timeline, library books, articles, and news (~290 documents) |
get_entry | One wiki/timeline/articles/news entry as markdown with claim type, editorial pass, typed related-content edges, and citation URLs; lang selects any of the 9 site languages |
get_passage | A library chapter as numbered paragraphs with stable reference IDs (e.g. GEN-1:26); ref filters to one verse |
get_source | Bibliography records by stable ID or search, with authority tier, source family, and stance — critical sources included |
compare_traditions | Rows from the comparative datasets: flood-myths, divine-council-index, theomachy-crossrefs, world-ages, prophets-and-religions |
query_graph | The typed 1-hop ego network of any corpus node (see_also, in_body, cites_source, same_tradition, same_age, comparison_of) |
Interpretation layer#
| Tool | Returns |
|---|---|
get_interpretation | The framework’s reading of a topic — curated context documents for framework topics, the closest wiki reading otherwise; always carries an interpretation-layer notice |
get_method | The editorial methodology: claim-type taxonomy, six-source rule, editorial passes |
Meta#
| Tool | Returns |
|---|---|
get_glossary_term | Core terminology with renderings in all 9 site languages; exact-id match wins |
Resources#
The server also exposes MCP resources for attachable context: the six curated narrative documents at woh://context/{overview, hypothesis, terminology, timeline, sources, method} and the API’s manifest at woh://llms.txt.
Discovery#
An agent that has not been handed the endpoint can find it. The server card is served from three places — all generated from the repository’s server.json, so they cannot drift:
| URL | Notes |
|---|---|
https://mcp.wheelofheaven.world/.well-known/mcp/server-card.json | Canonical |
https://mcp.wheelofheaven.world/.well-known/mcp.json | Same document, alternate convention |
https://www.wheelofheaven.world/.well-known/mcp/server-card.json | Copy on the apex site, where scanners look |
The nine tools are described independently at /.well-known/agent-skills/index.json, and both documents are referenced from the site’s API catalog. Every other path on the MCP host returns 404 — including /.well-known/oauth-*, deliberately, so that RFC 9728 discovery misses cleanly and clients fall back to anonymous access rather than attempting a token exchange against a server that has no auth.
In the browser, every page of www.wheelofheaven.world registers the same nine tools through WebMCP, proxied to this server over Streamable HTTP, alongside eight page-level tools (site search, page metadata and text, citations, navigation). An agentic browser therefore gets the corpus without being told the endpoint exists. The names and schemas are copied from tools/list, so a change to the tool set has to be mirrored there — see the maintenance table in Agent discovery.
Full detail: Agent discovery and crawler policy.
Architecture notes#
- Cloudflare Worker (Durable Object per session, Streamable HTTP), one TypeScript codebase; a stdio npm build shares the same tool registration.
- Stateless over the upstream API: no data of its own. The ~600 KB search corpus is cached in-memory per isolate for one hour, matching the API’s edge-cache TTL.
- If a tool ever needs data the API lacks, the fix is an additive
/v1/endpoint first, then a thin tool — the API remains the single public query surface.