LidLiftVercel + Cloudflare + OpenAI

Stop AI agents from choosing the wrong tool.

LidLift scores prompt-tool fit before execution, turns it into an explicit allow, review, clarify, or block decision, and exposes the same analyzer through a live API and remote MCP server.

Run the analyzer Open live API

Guardrail outcomes: allow, review, clarify, block

Live API surface on your own domain

Shared runtime surfaces: HTTP and MCP

Selected launch integration

The best fit for this product is a split runtime: Vercel for the operator console, Cloudflare for the remote MCP surface, and OpenAI Responses for the optional narrative layer.

Web control plane on Vercel

Fast Next.js App Router deploys, typed API routes, and a polished launch surface for analysts and operators.

Remote MCP server on Cloudflare

Streamable HTTP transport on `/mcp`, low-latency edge distribution, and a public endpoint that shares the same analysis engine.

OpenAI Responses for operator briefings

Deterministic ranking stays primary. Responses adds a concise launch-grade explanation when you want human-readable triage.

Interactive ConsoleVercel UI + Cloudflare MCP

Run the launch path before the model does

Score prompt-tool fit, inspect ranked alternatives, and optionally generate an operator brief with OpenAI Responses.

Live Preview

docs_search

User prompt

Generate operator brief with OpenAIWhen `OPENAI_API_KEY` is set, the API adds a short launch-grade narrative on top of the deterministic ranking.

5 candidate tools available for this run

Deterministic rankingShared with MCP server

The browser preview uses the same core ranking logic as the API and Cloudflare MCP endpoint, so your product UI and agent runtime stay aligned.

Best Fit

The current top-ranked candidate from the same engine used by the MCP server.

reviewmoderate

Best Candidate

docs_search

Alignment 46%

Dissonance 54%

Search developer documentation and return page content snippets.

Guardrail

review

Mismatch signals are strong enough that a person should review this tool choice before execution.

Gate this tool behind human review and surface a better-matched alternative.

Matched Intent

read

positiveOperation fit

Matched intent(s): read.

Ranking Board

Lower dissonance is better. Use this to gate write-capable tools before execution.

Rank 1

docs_search

reviewmoderate

Search developer documentation and return page content snippets.

Mismatch signals are strong enough that a person should review this tool choice before execution.

Dissonance: 54%

Alignment: 46%

Rank 2

weather_lookup

reviewmoderate

Read-only weather lookup for current conditions and short forecasts.

Mismatch signals are strong enough that a person should review this tool choice before execution.

Dissonance: 54%

Alignment: 46%

Rank 3

variant_lookup

reviewmoderate

Resolve archived genomic variants and species metadata from EVA.

Mismatch signals are strong enough that a person should review this tool choice before execution.

Dissonance: 56%

Alignment: 44%

Rank 4

tool_fit_analyzer

clarifyhigh

Analyze prompt-to-tool fit, detect mismatches, and rank candidate tools by dissonance.

The intended operation is not clearly supported by this tool, so execution should pause for clarification.

Dissonance: 81%

Alignment: 19%

Rank 5

github_publish

blockcritical

Create branches, commit changes, push code, and open pull requests on GitHub.

The prompt reads as inspection or retrieval, but this tool exposes a write path.

Dissonance: 100%

Alignment: 0%

Operator Brief

Optional narrative summary from OpenAI Responses. The ranking itself remains deterministic.

Set OPENAI_API_KEY to enable operator brief generation, then rerun the analyzer.

Block costly misfires

Read-vs-write mismatches do not just get a scary score. They can be hard-blocked before an agent publishes, refunds, deletes, or sends the wrong thing.

One engine, every surface

The browser preview, Next.js API, and Cloudflare MCP server all run the same analyzer, so demos, operators, and agents stay aligned.

Built to spread

A public API, a live MCP endpoint, and a sharp share card make this easy to demo, post, and reuse in other tool-routing stacks.

Launch Checklist

What is already productized

The remaining work is distribution, benchmarks, and operator adoption. The product surface, API contract, and MCP server are already aligned.

Shared `@tool-dissonance/core` package for heuristics, schemas, and sample catalog.

Next.js operator console with live preview, ranked results, and explicit guardrail decisions.

Cloudflare Worker exposing `analyze_tool_fit` and `rank_tools` over remote MCP transport on `/mcp`.

Health endpoints and environment examples for both the web app and the Worker.

Surface Area

Launch interfaces

The product ships as both a human-facing console and an agent-facing endpoint. That keeps operator workflows and agent workflows on the same scoring contract.

API map

Minimal endpoints needed to ship the first production release.

POST /api/analyze

Rank prompt-tool fit for the supplied catalog and optionally attach an OpenAI-generated operator brief.

GET /api/health

Expose web readiness and whether OpenAI narrative generation is configured.

POST /mcp

Remote MCP transport on Cloudflare Workers. Provides analysis tools to external clients over Streamable HTTP.