Proof of Usefulness Reports

urweb — AI-Powered Personalized Web Audit for SMBs

urweb is a fully autonomous B2B web audit service that discovers real SMB prospects, diagnoses their current website, designs bespoke CSS/HTML mockups tailored to their branding, and sends personalized email previews — with 13.5% CTR and 7 real business replies from 297 delivered emails.

agenttrace: cost and latency tracking for agent runs

A tiny zero-dep JavaScript library that groups LLM calls into runs and reports total cost, p50 and p95 latency, and a by-model breakdown. Composes with prompt-cache observability and per-call cost calculators so you get a per-run report instead of a per-call spreadsheet.

agentvet: validate tool args before execution

A tiny zero-dep JavaScript library that wraps each AI agent tool function with arg validation. If the model passes invalid arguments, agentvet throws a ToolArgError with an LLM-friendly retry hint so the model can self-correct. Five lines to wire into existing tool definitions; works with any agent framework.

agentcast: structured output for any LLM call

A tiny zero-dep JavaScript library that turns any LLM call into a typed-output call. Pass a Zod schema, get validated typed data back, or get a clear retry-with-feedback loop until the model returns valid output. BYO LLM via async closure; works with Anthropic, OpenAI, Bedrock, anything that takes a prompt.

agentsnap: snapshot tests for AI agents

A tiny zero-dep JavaScript library that records AI agent tool-call traces and snapshots them like Jest. Run your agent, snapshot the tool sequence, set baselines, fail CI when an agent silently changes which tools it calls. The kind of regression test most agent stacks ship without because nobody wants to set up the harness.

agentguard: network egress firewall for AI agents

A tiny zero-dep JavaScript library that gives every AI agent tool a declarative allowlist of domains it can fetch from. If a tool tries to reach a domain you have not allowed, agentguard throws. Designed for teams running agents that browse the web or call third-party APIs and want a hard guardrail rather than a hope.

bedrock-kit: opinionated AWS Bedrock client wrapper

A small, opinionated AWS Bedrock client wrapper that handles three things native SDKs leave to you: adaptive throttle (back off when Bedrock rate-limits, recover smoothly), cache-aware cost tracking (per-call dollars based on actual cache hits), and structured-output parse-and-repair (when the model returns broken JSON, fix it locally before retrying). Single-cloud, single-purpose, MIT.

cachebench: prompt-cache observability for LLM APIs

A Python library that turns prompt-cache hits/misses into a numeric report you can act on. Per-call hit ratios, dollars saved, regression alerts when cache effectiveness drops, and miss-aware retry helpers. Works across Anthropic, OpenAI, and Bedrock so teams running multi-provider LLM pipelines can finally see what their cache layer is actually saving them.

driftvane: composable RAG and agent drift detectors

A Python library that composes embedding, retrieval, response, and latency drift detectors into one DriftReport for RAG and agent systems. Library-only, zero server, zero UI; intended to live next to your existing search and retrieval stack and surface when answer quality is silently degrading. Pairs naturally with intelligent search backends so teams can see when retrieval starts feeling stale before users complain.

agentmemory: pull-model episodic memory for AI agents

Pull-model episodic memory for AI agents with real deletes and an audit trace. Ships as a tiny zero-dep npm package plus a Python sibling, with a Hermes Agent plugin (hermes-agentmemory) that wires it as data infrastructure for the open-source 151k-star Hermes Agent. Built for teams that need GDPR-style real deletes and a tail-able audit log instead of background consolidation that bakes memories you cannot un-bake.

Outdo

Outdo is a real-time competitive productivity app where friends race to complete their daily tasks on a live leaderboard. Users add tasks with time estimates, join a shared room with a code, and compete. Their progress is weighted by task duration so a 2-hour task counts more than a 5-minute one. It was built with React, Firebase, and Claude AI for task extraction, it's live and being used by real people today.

RoastRocket

RoastRocket stress-tests startup ideas in 10 minutes. 15 questions across 5 perspectives — founder fit, market reality, timing, business model, and moat. You get a score and a verdict: build, pivot, or kill. Free, no signup needed.