Model Context Protocol
Built for MCP, on the standard.
The graph is built to be read over the Model Context Protocol: one server, typed tools over entities, relationships and taxonomies. Client access is not open yet — the transport in front of it is still being finished, and we would rather say so than have you spend an afternoon on a connection that cannot authenticate. Spec, capabilities and the tool surface are published below for procurement review.
The 6 servers
One server. Nineteen graph tools.
One server exposing the typed graph — entities, relationships, taxonomies, and the evidence behind each answer. Listed here so you can review the surface before access opens.
| Server | Purpose | Tools |
|---|---|---|
product-graph | Query the typed product graph from any MCP client — entities, relationships, taxonomies, and the evidence behind them. | 19 |
Tested with these clients
Plug into whatever AI client your team already uses.
Claude Desktop
Add via:
Settings → Connectors → Add MCP ServerClaude Code
Add via:
.mcp.json or `claude mcp add`ChatGPT Enterprise
Add via:
Admin → Connectors → Custom MCPCursor
Add via:
.cursor/mcp.jsonWindsurf
Add via:
Cascade Settings → MCP ServersVS Code (Copilot Chat)
Add via:
settings.json `github.copilot.chat.mcp.servers`Continue.dev
Add via:
config.yaml mcpServers sectionOpenAI Agents SDK
Add via:
agents.mcp.MCPServerHttp(...)Technical specification
The procurement-questionnaire answer in one table.
| Field | Value |
|---|---|
| MCP spec version | 2025-11-25 (stable) · 2026-07-28 release candidate (preview) |
| Transports | stdio · HTTP+SSE · WebSocket (notifications) |
| Authentication | OAuth 2.1 with PKCE · service tokens for headless clients |
| Tenant isolation | Per-token scoping; tenant graph slices enforced at the server layer |
| Rate limits | Per tier · Starter 100 req/min, Growth 500 req/min, Enterprise custom |
| Audit log | Every tool call logged with caller, tool, args hash, latency, result digest |
| Citation passthrough | Every Resource response carries the same provenance schema as the workspace UI |
| Protocol compliance | Tested against MCP Inspector + reference clients in CI on every release |
Why MCP, not bespoke
The Linux Foundation calls MCP the USB-C of AI integrations.
Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation in December 2025. Google, Microsoft, AWS, OpenAI, Block, and Cloudflare back it. Building a bespoke integration today means catching up to the standard tomorrow · we shipped on the standard from day one.
Wire APEX into Claude Desktop in five minutes.
We'll send you the OAuth credentials and walk you through the .mcp.json on the call. You'll be querying the graph from inside your existing AI client by the end of the demo.