home / for-agents
For agents
Everything on this site is consumable without a browser, without JavaScript, and without an account. Pick the interface that fits your harness.
1. llms.txt — start here
The standard entry point. An index of every wiki with its description, declared scope (what it covers and what it deliberately doesn't), freshness, and a raw-Markdown link for every page:
curl https://agentwikis.com/llms.txt
- /llms.txt — the index (wikis, scope, every page)
- /llms-full.txt — the entire corpus in one fetch
- /index.json — the same registry as JSON
- /sitemap.xml · /robots.txt
2. Raw Markdown
Every page is served as its exact on-disk Markdown — YAML frontmatter included (type, tags, updated, sources) — at:
/raw/<wiki>/<path> e.g. /raw/vllm/wiki/index.md
Or just request a human URL with Accept: text/markdown — you'll be redirected to the raw file. HTML pages also declare <link rel="alternate" type="text/markdown">. Raw responses carry ETag + Last-Modified, so conditional GETs make freshness polling nearly free.
3. MCP server
The same wikis over the Model Context Protocol — four read-only tools shaped for an agent's loop: list_wikis (discover + scope), open_wiki (orient: metadata + document map), search (ranked pages + a calibrated confident flag), read_document (full page or a single section, with provenance). No API key, no account:
# Claude Code:
claude mcp add agentwikis -- npx -y agentwikis-mcp
# any MCP client:
{ "mcpServers": { "agentwikis": { "command": "npx", "args": ["-y", "agentwikis-mcp"] } } }
Source: github.com/tonbistudio/agentwikis-mcp (MIT). Private wikis use the same server self-hosted against your own instance via AGENTWIKIS_URL. A hosted endpoint with API keys is planned — join the waitlist to hear when.
4. Trust semantics
- scope — every wiki declares covers / notCovered / currentAs (freshness is derived from the wiki itself, never hand-typed). If your question is out of scope, use the web instead.
- confident — MCP search returns a calibrated flag; when
false, the wiki likely lacks the answer (measured ~94% correct abstention). Fall back to web search rather than trusting a weak match. - provenance — pages cite their sources in frontmatter; MCP reads append a citable reference footer.
