How it works
A seven-stage pipeline, from a brief to a shortlist.
The research agent runs autonomously from a finalized brief. Scoring arithmetic happens in code, not in the model — so the numbers are reproducible and every one of them cites its evidence.
- 01Discover
Generate diverse search strategies, build a longlist of 15–40 candidates, and probe each domain for a TinyRFP endpoint, llms.txt, or MCP card.
- 02Screen
Apply disqualifiers using the cheapest evidence first — geography, license, category fit — logging every elimination with its reason and source.
- 03Investigate
Gather evidence per rubric criterion from the vendor's site, third-party reviews, news and regulatory records, financial signals, and community discussion.
- 04Engage
Interview provider endpoints for what the web couldn't answer — respecting your disclosure tier, never sending your identity by default.
- 05Verify
Cross-check high-weight claims against authoritative registries: licenses, FDA clearances, ISO registrations, status pages. Contradictions are surfaced prominently.
- 06Score
Score each candidate per criterion on an anchored 0–5 scale, citing specific evidence. Vendor claims alone can't max out a criterion that requires third-party proof.
- 07Report
Produce the shortlist report — recommendations, a comparison matrix, candidate dossiers, methodology, and full appendices — in Markdown, HTML, and PDF.
Discovery
No central gatekeeper
Discovery works in layers, and the order matters. First, an open web search — which requires nothing from any provider. Then, per-domain probing of /.well-known/tinyrfp.json and llms.txt on the candidate domains that search turned up. An optional community index can come later for direct lookup.
This ordering is why fairness is structural, not a policy promise: a provider with no endpoint is found and evaluated exactly the same way. An endpoint adds fidelity and interactivity — it is never a toll booth.
Evidence
Every claim carries its provenance
The atomic unit of the whole system is the evidence record: a claim, its source, when it was retrieved, a verification outcome, and a provenance class.
- Independently verifiedConfirmed against an authoritative registry or primary source.
- Third-partyPublished by someone other than the provider — reviews, news, analysts.
- Vendor-publishedOn the provider's own public website or materials.
- Vendor-claimedStated interactively via an endpoint or contact — unaudited, per-session.
The capping rule
Scope
Where the agent stops
TinyRFP automates the top of the funnel — from identifying a need through a defensible shortlist. It is deliberately not a closer.
There is no negotiation, no contracting, no e-signature, and no payment. The agent recommends and ranks; a human makes the decision. For finalists, the report turns the questions the research couldn't answer into a ready-to-use demo and call list — so the human conversation starts already informed.
The tools are coming soon.
TinyRFP ships as portable skills and a reference MCP server you run inside the AI assistants you already use — no new app to install.