Finding the Right MCP Tool

There are about thirty MCP tools across the two servers this wiki exposes. Most agents only ever need five of them. This runbook is the decision tree.

When to use this runbook

Any time you catch yourself reaching for get_page when you really only needed orientation, or trying to write through /knowledge-mcp (which is read-only). Re-read this when you start a new agent session.

Context

Two MCP endpoints, partitioned by trust level:

• /knowledge-mcp — read-only. Hosts retrieval (search_knowledge, retrieve_context, find_similar), navigation (list_clusters, list_tags, list_pages_by_filter, get_page_by_id, traverse_relations), knowledge-graph traversal (discover_schema, query_nodes, get_node, traverse), and the agent-grade projection (get_page_for_agent).
• /wikantik-admin-mcp — writes + analytics. Hosts page edits (write_pages, update_page, rename_page, delete_pages), verification (mark_page_verified, verify_pages), structural analytics (get_backlinks, get_page_history, diff_page, get_outbound_links, get_broken_links, get_orphaned_pages, get_wiki_stats, preview_structured_data, ping_search_engines), and graph proposals (propose_knowledge, list_proposals).

Walkthrough

Two questions disambiguate every scenario:

1. Read or write? Writes go through /wikantik-admin-mcp only and require an author-configurable agent. If you're not certain you have write privileges, default to read.
2. Full body or just the shape? If you only need to cite the page or orient yourself, get_page_for_agent returns the token-budgeted projection — verification state, key facts, headings, typed relations. If you need the prose, fall back to get_page.

For the broader retrieval question — "which tool finds me a page" — defer to ChoosingARetrievalMode.

Pitfalls

The frontmatter pitfalls capture the recurring mistakes. The most expensive in tokens-per-error is the get_page-for-orientation antipattern: agents pull the entire markdown body when a 4 KB projection would have sufficed.