A Live Knowledge Graph of Your Site
Pages and sections become typed nodes, links become edges with provenance. Agents call graph_node, graph_neighbors, graph_search, and graph_path instead of scraping HTML.
Your site opens its own door to AI agents
Grapheway turns your website into a live, typed knowledge graph that AI agents discover, traverse, search, and act on. One package, zero dependencies, no paid crawlers. Your site opens its own door.
Real commands and real output. Every snippet below is the exact format the project writes to disk.
$ bunx grapheway probe https://expressjs.com
Probed Express.js (https://expressjs.com)
pages: 25
headings: 521
edges: 1927 (1361 extracted, 566 inferred)
api: 12 endpoints (https://expressjs.com/openapi.json)
Serving "Express.js" as an agent surface:
discovery http://localhost:4321/.well-known/agent
graph http://localhost:4321/graph/v1
events http://localhost:4321/graph/v1/events (realtime SSE)
MCP http://localhost:4321/mcp
# That legacy site is now an MCP server. Point Claude Desktop
# or Cursor at localhost:4321/mcp and its docs become tools.Every capability between install and a working setup. Search by keyword, filter by category.
Pages and sections become typed nodes, links become edges with provenance. Agents call graph_node, graph_neighbors, graph_search, and graph_path instead of scraping HTML.
The /mcp endpoint is a spec-aligned Model Context Protocol server. Claude Desktop, Cursor, VS Code, and any MCP client get your graph tools and actions as native tools, plus pages as markdown resources. No key, no account.
Point grapheway probe at any URL, no site involvement needed. It extracts the site's own knowledge, title, navigation, headings, links, and OpenAPI endpoints, then serves the same agent surface locally or exports it as JSON.
An A2A-style agent card tells any agent exactly what your site exposes, how, and where. Discovery is a URL, not a partnership negotiation.
Your basic actions, like check status or query, become callable tools over HTTP and MCP. Same implementation, both surfaces.
grapheway/compat optionally serves llms.txt, agents.txt, robots.txt, and sitemap.xml at runtime, for the agents that still probe them. Fully decoupled: mount it or do not.
The core is framework agnostic and dependency free. It runs in Node, Bun, and Deno, works with Express, Hono, Next.js, plain node:http, or static hosts. The handler interface is four fields in, four fields out.
Real numbers from running grapheway probe against expressjs.com with default settings. Your numbers will differ per site.
| Surface | What agents got | Provenance | Freshness |
|---|---|---|---|
| grapheway probeLEADING | 25 pages, 1,927 edges | extracted + inferred | live, re-crawl on demand |
| Scraping HTML | 25 downloads per agent | markup, untyped | stale on next change |
| llms.txt snapshot | one flat file | hand maintained | static until edited |
| Paid crawler API | same pages | third party | on their schedule |
Six subpath exports, about 6,000 lines of TypeScript, zero dependencies. Where the code lives.
1,336 lines. The graph model: typed nodes, typed edges, properties, versioning. Graph operations, live patching with structural validation, diff between snapshots, the discovery manifest, and Schema.org JSON-LD serialization.
2,154 lines. The drop-in agent endpoint: createGrapheway, 14 routes across discovery, graph, viewer, agent API, and MCP. Framework adapters for node:http, Express, and Hono. Live SSE events, patchGraph, and GEO head injection.
985 lines. The agent-side crawler: fetches any site same-origin only, respects robots.txt, extracts nav, headings, links, and OpenAPI specs into tagged edges, then serves or exports the result.
540 lines. The optional legacy module: llms.txt, agents.txt, agents.json, robots.txt, and sitemap.xml served at runtime. Fully decoupled from the core.
385 lines. The typed GraphewayClient for agents: discovery, graph summary, node, edges, search, path, native traverse, manifest, pages as markdown, and action calls.
661 lines. The grapheway command: probe, gateway, serve, mcp-config, audit, and generate. mcp-config prints the exact client snippet to paste.
Grapheway is GPL-3.0 and built in the open, with a white paper and a public changelog. If your agents stopped scraping today, star the repo or pass it to another web team.