Connect your agent to the graph
The Knowledge Engine runs a remote MCP server — point Claude Code, Claude Desktop, Cursor, or any MCP client at it and your agent can resolve a company, traverse who supplies and competes with whom, pull the dated evidence behind every claim, and ask what was knowable on any past date. Every answer comes back as sourced JSON with citations, not guesses.
Setup takes about a minute. Copy the config for your client below, drop in your key, and ask “Who supplies Nvidia?”
1. Your connection details
This key is shared across the team. Treat it like a password — it grants read access to the full graph. It can be rotated at any time; if a config stops working, grab the current key from this page again. Sign in to reveal it.
2. Add it to your client
Claude Code (recommended)
One command in your terminal — no files to edit:
claude mcp add --transport http knowledge-engine https://knowledge-engine.ai/mcp \ --header "Authorization: Bearer <sign in to reveal your team API key>"
Then run /mcp inside Claude Code to confirm knowledge-engine is connected. Add --scope project to share the config with a repo via .mcp.json (omit the key there and use an env var if the repo is shared).
Cursor
Add to ~/.cursor/mcp.json (global) or .cursor/mcp.json (per-project):
{
"mcpServers": {
"knowledge-engine": {
"url": "https://knowledge-engine.ai/mcp",
"headers": { "Authorization": "Bearer <sign in to reveal your team API key>" }
}
}
}Claude Desktop
Claude Desktop needs the mcp-remote bridge to send the auth header. Add this to your claude_desktop_config.json (Settings → Developer → Edit Config), then restart the app:
{
"mcpServers": {
"knowledge-engine": {
"command": "npx",
"args": [
"mcp-remote",
"https://knowledge-engine.ai/mcp",
"--header",
"Authorization: Bearer <sign in to reveal your team API key>"
]
}
}
}Anything else (raw)
Any MCP client that speaks Streamable HTTP works — point it at the endpoint and send Authorization: Bearer <key>. Quick smoke test from a shell:
curl -s https://knowledge-engine.ai/mcp \
-H "Authorization: Bearer <sign in to reveal your team API key>" \
-H "Content-Type: application/json" \
-H "Accept: application/json, text/event-stream" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'3. Try it
Once connected, just ask your agent in natural language — it will pick the right tools:
- “Who supplies Nvidia, and who competes with it?”
- “What’s the latest sourced evidence on Samsung’s HBM4 qualification?”
- “List the companies in the HBM segment.”
- “What did we know about CoWoS capacity as of March 2025?”
Eight ways to query the graph
The server exposes eight governed tools. Each returns compact, source-tagged rows — never raw SQL or a black box — and degrades to an honest empty result rather than guessing.
Map the names in a question to the canonical companies we track, with each one’s place in the AI-stack taxonomy. The anchor for everything else — call it first.
“Resolve ‘SK Hynix’, ‘Vera Rubin’.”
Every dated, attributed claim we hold on a company — publisher, date, stance, and an analytical-quality score on each one.
“What’s the latest on Samsung’s HBM4 yield?”
The typed supply-chain and competitive edges — SUPPLIES_TO, PARTNERED_WITH, COMPETES_WITH — straight from the graph. Multi-hop by following a counterparty.
“Who supplies Broadcom? Who competes with it?”
Structured quantitative metrics for a topic or measure — capacity, funding, throughput — each with its subject, basis, date, and source.
“How has HBM4 contract pricing moved?”
Full-text search across the source corpus for fuzzy questions or names that don’t resolve cleanly — the fallback that keeps the agent honest about coverage.
“Anything on co-packaged optics test yield?”
Bitemporal replay: only the claims recorded on or before a past date. Ask what was knowable then, and compare it to now.
“What did we know about Samsung as of March 2025?”
The complete, deterministic roster of companies in a stack segment, sub-segment, or layer — the full set, not a partial sample. Best for “who are the players in X”.
“List every company in the HBM segment.”
Companies ranked by how often a theme shows up in their sourced claims — a fast way to find who matters on a given topic.
“Which companies are most tied to CoWoS?”
An event-sourced, bitemporal graph
- Typed events, not free text. Claims are extracted into a fixed ontology — qualifications, supply agreements, capacity moves, funding — so the graph is queryable by kind of event.
- Every claim carries provenance. Source, publisher, date, a verbatim evidence quote, and a credibility tier — so an answer can always be traced back.
- Two clocks. When something happened vs. when we learned it — which is what makes “as of” questions answerable instead of hand-wavy.
- Companies placed in one stack. Every entity sits in a shared AI-infrastructure taxonomy, so cross-topic questions join cleanly.
The graph answers only from what it holds— it cites the source and date for what it asserts, and when the graph is silent it says so rather than inventing a supply chain. In a product whose whole value is trust, a confident wrong answer is worse than an honest “we don’t track that yet.”
No — once the config is in your client, it just works. Come back here only to grab the key again (e.g. after a rotation) or to copy a config for a new machine.
The AI-infrastructure supply chain — companies, typed relationships, sourced claims, and quantitative metrics, all placed in one stack taxonomy. Curated and auto-discovered sources (analyst newsletters, trade press, primary filings, conference papers) are extracted into typed claims before they enter the graph.
Yes. Every tool is a governed read over the graph — nothing an agent does through MCP can modify the data.
Use Ask the Graph — the same tools, run as a hosted chat that returns a sourced answer with a faithfulness verdict and a sources panel.