Connect to this knowledge graph over MCP
This site's knowledge graph exposes a read-only MCP server. Any MCP client that speaks streamable HTTP can connect, no API key required. It only serves public-tier content: your agent can search concepts, fetch a single concept page, pull the whole graph, and ask grounded questions.
Endpoint and protocol
POST https://api.paulkuo.tw/mcpJSON-RPC 2.0 over POST only (other methods get 405), stateless streamable HTTP, no batch requests, no authentication. Browser pages are restricted by an Origin allowlist, so call it from an agent or server side.
Setup
One line in Claude Code:
claude mcp add --transport http paulkuo-wiki https://api.paulkuo.tw/mcpOr try it with curl first:
curl -s https://api.paulkuo.tw/mcp \
-H 'content-type: application/json' \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"wiki_search","arguments":{"query":"AI"}}}'Any other MCP client that supports streamable HTTP just needs the endpoint above.
The four tools
| Tool | What it does | Scope and limits |
|---|---|---|
wiki_search | Keyword search over public concept pages (matches titles, tags and excerpts); returns matching pages with excerpts. | Indexes public concept pages only; source material and full article bodies are not included. |
wiki_get_concept | Fetches one public concept page in full (frontmatter plus body) by slug, e.g. from a wiki_search result or a [[slug]] citation in a wiki_ask answer. | Slugs are lowercase letters, digits and hyphens only; unknown or non-public slugs get a 404-style error. |
wiki_graph | Pulls the whole public knowledge graph: nodes (id / title / type / pillar) and undirected edges. | No parameters; returns the full graph in one response, which is large. |
wiki_ask | Grounded Q&A: retrieves public wiki content first, then answers with [[slug]] citations. | Questions up to 500 characters; refuses to answer when retrieval finds no grounding. The only tool that burns AI inference cost, so it has daily quotas. |
Visibility red line
All four tools read the public tier only. Internal notes, unpublished material, and anything that stays private even after de-identification are simply not in this server's world.
Rate limits
Roughly 30 requests per minute on /mcp, 10 tool calls per minute, and for wiki_ask an extra 5 per minute plus 30 per IP per day; a site-wide daily budget returns 503 when exhausted. Per-minute limits are best-effort in-memory counters; the daily quotas and site-wide budget are stored counters that expire after 24 hours. Numbers may change.
Question logging
Questions sent to wiki_ask are logged to study question patterns and improve retrieval, and are auto-deleted after 90 days. Before logging, emails, Taiwanese mobile and landline numbers, Taiwanese national ID patterns and digit runs of 9 or more are masked by common-format matching; other formats or other personal data may not be masked, so please keep them out of your questions. The question log itself contains no IP or browser info; a separate daily-quota counter is keyed by IP and expires after 24 hours.
Answer language
Wiki content and wiki_ask answers are primarily in Traditional Chinese. wiki_search matches titles and tags in that zh-TW corpus, so Chinese keywords work best; an English term only matches when it appears verbatim in a title or tag (e.g. AI, MCP). For cross-language questions, use wiki_ask instead.
The /ask alias
The same Q&A capability is also exposed as an NLWeb-style HTTP alias (GET or POST https://api.paulkuo.tw/ask) for integrations that skip MCP. "NLWeb-style" describes the shape, not verified spec compliance.
Not the same as the in-page WebMCP
Site pages also ship client-side WebMCP tools (/.well-known/mcp.json) that run in the browser. This page is about the server-side MCP endpoint; the two are independent.