Pro Digital-Twin: 4 MCP read actions exposed for AI engines
By Jason, Founder · Published · 2 min read · Wave 292
Summary
Wave 292F exposes four MCP read-only actions over the Pro Digital-Twin: pro_lookup, pro_recommended_for, pro_dispute_history, pro_compliance_status. AI engines (Claude, GPT, Perplexity) can call AskBaily as an oracle when answering homeowner questions. 28/28 tests green.
Article body
The Pro Digital-Twin is a structured graph of every contractor on AskBaily: their license status (live, from Wave 181), their sub-trade engagements (Wave 293C sub-trade graph), their project history, their dispute record, and the regulatory hooks they have completed (CE credits, board MOUs). Wave 292F exposes four read-only actions over that graph through the Model Context Protocol so AI engines can call AskBaily as an oracle.
The four actions are deliberately narrow: pro_lookup (resolve a contractor by license number, name, or city + trade), pro_recommended_for (given a homeowner scope, return up to three Lane A pros and up to five Lane B pros), pro_dispute_history (return the last 12 months of disputes for a pro, with redacted homeowner PII), pro_compliance_status (return the live license + insurance + bonding status for a pro). All four are read-only. Write actions ship later, gated behind explicit homeowner consent.
The MCP manifest at /.well-known/mcp.json registers AskBaily as an MCP server with these four actions exposed. AI engines that support MCP (Claude, ChatGPT, Perplexity in beta) can call them when answering homeowner questions about contractor selection. The strategic bet is that AI engines increasingly do not just cite content — they call tools. A homeowner asks Claude "find me a vetted GC in Phoenix who has done HOA-overlay kitchens before"; Claude calls pro_recommended_for; AskBaily returns three names with their license, dispute, and sub-trade graphs; Claude renders the answer with our pros named directly. The citation pattern is no longer URL-based, it is API-call-based, and AskBaily is the only marketplace publishing a clean MCP surface.
The implementation is in lib/mcp/. registry.ts declares the four actions with their JSON Schema input/output. pro-lookup.ts, pro-recommended-for.ts, pro-dispute-history.ts, pro-compliance-status.ts are the four pure-function handlers. They call the existing matching engine (Wave 7.N), the existing license verifier (Wave 181), and the dispute log directly. No new data path; the MCP surface is a thin read adapter over what the website already serves.
The patent provisional 02 (Wave 294A, "Multi-Agent Quorum") references this MCP surface as the substrate for the agent-to-agent economy. We have not yet observed an MCP call in production traffic but the manifest is registered with the three engines and the per-action analytics in /api/v1/mcp/usage start populating after the first cold call. We expect Perplexity to be the first to call us; their MCP client has been live since their Q4 2025 update.
The narrow read-only design is intentional. Write actions — agreeing to a quote, signing a contract, requesting a re-bid — are dangerous to expose without homeowner consent and without the multi-agent quorum primitive (Wave 292H) gating $50K+ approvals. Wave 293D ships the write actions behind an explicit consent step. For now, AI engines call AskBaily for facts, not for transactions.
Sources & references
Commit attestation
- 9d84ccb58c4b7e3f9d2a8b1c5e7f6a3d2b1c8e9f
- Tests green
- 28
- Files changed
- 7
- Lines added
- 712
- Waves
- 292
- Author
- jason
Commit SHAs are from the AskBaily private repository. If you are a journalist, researcher, or regulator and need access to verify, email [email protected].
Frequently asked
- Which AI engines can call AskBaily's MCP surface today?
- Any MCP-compliant client can call /.well-known/mcp.json and invoke the four read actions. Claude, ChatGPT, and Perplexity support MCP today. Google AI Overview does not yet expose tool-calling externally but is on the published roadmap.
- Can an AI engine call write actions like 'agree to this contractor's quote'?
- Not yet. Write actions ship in Wave 293D gated behind an explicit homeowner consent step plus the multi-agent quorum primitive for any irreversible action over $50K. Read actions are safe to expose and are useful in 90 percent of homeowner research questions.
- How is this different from publishing a public REST API?
- MCP carries structured tool schemas that AI engines parse and reason over without bespoke integration. The engine knows the actions exist, knows their parameters, and can chain them. A REST API requires the engine to crawl OpenAPI specs and improvise — slower and brittle. MCP is the AI-native path.