agentwikis-mcp

MCP server for Agent Wikis — maintained, agent-readable knowledge bases compiled from messy, fast-moving sources. Read-only, no API key, no account.

Four tools shaped for an agent's loop:

Tool	What it does
list_wikis	Every wiki + its declared scope (covers / not covered / current-as-of)
open_wiki	One wiki's metadata + full document map — orient in a single call
search	Ranked pages with a calibrated confident flag — when false, fall back to web search
read_document	Full page or a single section, with a provenance footer

The scope declarations and the confident flag exist so your agent knows when not to trust the wiki — measured ~94% correct abstention. Details and the full eval: agentwikis.com/why-wikis.

Install

Claude Code

claude mcp add agentwikis -- npx -y agentwikis-mcp

Codex (~/.codex/config.toml)

[mcp_servers.agentwikis]
command = "npx"
args = ["-y", "agentwikis-mcp"]

Any MCP client (JSON config)

{ "mcpServers": { "agentwikis": { "command": "npx", "args": ["-y", "agentwikis-mcp"] } } }

Self-hosted instances

Point the bridge at your own Agent Wikis deployment (e.g. a private-wiki install):

AGENTWIKIS_URL=https://wikis.internal.example npx -y agentwikis-mcp

What's inside

The bridge talks to agentwikis.com/api/* and returns the same payloads as the self-hosted stdio server — same document formatting, same provenance footers, same search calibration. Requires Node 18+.

Agent Wikis also serves everything as plain Markdown + llms.txt if you'd rather skip MCP entirely: see agentwikis.com/for-agents.

License

MIT