Quotrr for AI agents

Why we built this

Homeowners are starting to ask agents what a job should cost and who they should hire. Quotrr holds the answer for pool and landscape work in Sacramento: anonymized local price ranges, public contractor profiles with a Quotrr Score and verified Props. Rather than make agents scrape us, we publish a clean, identified, rate-limited surface so the data shows up correctly, with attribution, and so any action taken for a person carries that person's consent.

Open standards

Everything an agent needs to discover and read Quotrr is public:

• /llms.txt at the domain root. A plain-text map of the surface for language models.
• /.well-known/agents.json. The agent discovery file, served at the well-known path.
• mcp.quotrr.com. An MCP server preview, so MCP-capable agents can call Quotrr tools directly.
• api.quotrr.com/openapi.json. The public OpenAPI 3.1 spec for the full API.
• /abuse. How to report misuse of the surface and what we do about it.

Identity and consent: OAuth 2.0 with PKCE

Agent identity and human consent both run on OAuth 2.0 with PKCE.

• Agent identity. Agents that establish a verified identity get higher limits, and large agent platforms can apply for known-agent status. Unknown agents get a low anonymous tier: cached responses and a Turnstile-equivalent challenge until they establish identity.
• Human consent. Any action an agent takes for a person needs a delegated, scoped, expiring token signed by that human. The token names exactly what the agent may do, expires on its own, and can be revoked. Every consented action is logged in the human's audit trail and sensitive actions are reversible for 24 hours. After that window a published Prop follows the same rule as any other, removable only for fraud or a policy violation.

What is free, what needs consent, what is off-limits

Rate and cost limits

Limits are per agent identity. Known and verified agents get higher ceilings. Each identity also has a daily cost budget. When an agent breaches the budget, access auto-disables for the rest of the day. Every query is logged with identity, route, cost, and response size, so the limits are enforced against a real record, not a guess.

MCP tools

The MCP server preview at mcp.quotrr.com exposes these tools:

• get_price_range(trade, zip, scope). Anonymized local price range for a trade in a ZIP, at the requested scope.
• find_contractors(trade, zip, filters). Public contractor profiles matching a trade and ZIP, with optional filters.
• check_contractor_reputation(handle). The Quotrr Score and verified Props for a public contractor handle.
• request_quote(homeowner_info, scope, consent_proof). Request a quote for a homeowner. Requires a consent token. Logged and reversible for 24 hours.
• submit_review(verified_job_id, rating, props). Submit a review against a verified completed job. Requires a consent token. Logged and reversible for 24 hours.

The boundary still applies

Quotrr runs on Cloudflare D1 and R2 at the edge. Each contractor is a tenant with a privacy firewall. That same boundary applies to agents: no agent crosses a contractor tenant boundary without that contractor's grant. The agent surface is a window onto anonymized and published data, not a key to the building. See Security for the full model.