Developer access

Public Guard data API and MCP

These surfaces are read-only projections of reviewed public HOL Guard records. They do not expose workspace data, local prompts, source code, findings, secrets, or private telemetry.

HTTP JSON

JSON responses use versioned schemas, source-derived Last-Modified timestamps, strong ETags, conditional 304 responses, shared cache controls, and bounded application-level rate limiting.

GET /guard/facts.json
Current publishable product facts with evidence, verification date, expiry, and limitations.
GET /guard/security/campaigns/data.json
Current reviewed threat-campaign collection.
GET /guard/security/coverage/data.json
Current harness coverage assertions and evidence refs.
GET /guard/security/policies.json
Defensive policy guidance attached to reviewed campaigns.
GET /guard/research/ai-coding-agent-runtime-security-benchmark/data.json
Versioned fixture benchmark results and methodology.

Public MCP

Endpoint: POST /api/guard/public-mcp. Protocol: 2026-07-28. It is stateless and exposes only the same reviewed public datasets available over HTTP.

  • • Send MCP-Protocol-Version: 2026-07-28 and a matching Mcp-Method header.
  • • For tools/call, also send Mcp-Name matching the requested tool.
  • • Every tool result includes evidence IDs/URLs and material limitations.
  • • Request bodies are size-bounded and the endpoint is rate-limited.

Discovery limitation

AI assistants do not automatically discover, connect to, or invoke this MCP server merely because it is published here or referenced in llms.txt. A user, client, host application, or administrator must explicitly configure or connect the MCP endpoint. HTTP pages and feeds remain the primary broadly crawlable public interface.

Evidence and correction rules

A tool response is not an endorsement or an independent validation. Follow the returned evidence URL, verification date, scope, and limitations. Corrections are tracked at /guard/methodology/corrections.