Featured
- ★herdr-fleetA terminal dashboard, written in Rust, for a fleet of Claude Code agents running in herdr panes. One screen shows each agent's status, model, context use, estimated spend, time in its current state, last line of output, and flags for agents that look stuck or are looping on the same command.
Agent skills
Seven skills I use daily, published in my GitHub profile repository. Each is a folder with a SKILL.md plus the scripts and notes it relies on, and each runs in Claude Code, Codex and the ChatGPT desktop app (install instructions).
- S1shipA closed-loop acceptance gate. It builds the app, has a fresh verifier use it for real, fixes from the repro and re-verifies, until it passes or provably stops converging.
- S2cloud-browser-loginLogs into my own accounts on any website from a cloud session, in a real Chromium browser. It picks the safest login route, asks once for exactly the secrets it needs, and records what it learns about each site.
- S3retell-voice-agentDesign, build, audit and verify Retell AI voice and SMS agents, with a three-layer verification method and a catalogue of pitfalls found on real calls.
- S4deepgram-voice-agentBuild and gauntlet-test Deepgram Voice Agent API agents against persona scenarios, checking the real bookings and emails rather than the chat.
- S5dialoraBuild, configure and verify Dialora.ai voice agents end to end, from prompts and knowledge bases to tools, phone numbers and agency sub-accounts.
- S6autonomous-systems-auditA read-only health audit of a self-driving agent harness, its test suite, knowledge graph and memory stores, backed by a telemetry probe and file and line citations.
- S7pme-cdpDrives an already-logged-in Chrome over CDP instead of a fresh browser that gets bot-blocked and logged out.
Applications
Thirteen demo applications, each in its own directory of the source repository and hosted on Netlify. Each calls models through OpenRouter. The front ends are React and TypeScript built with Vite, and the back ends are Netlify Functions.
- AgentFlow
- DocMind
- CodeLens AI
- VoxAI
- DataPilot
- ModelArena
- ContentForge
- VisionLab
- InsightHub
- BrowseBot
- GraphScout
- GraphGate
- GraphSwarm
- 01Live
AgentFlow
Multi-Agent Research Orchestrator
Answers a research question from live Wikipedia and Hacker News sources. Four model agents run in a fixed order, then an audit checks every cited sentence of the report against its source and marks it supported, partly supported or not supported.
- 02Live
DocMind
RAG Document Intelligence
Ask questions about a live Wikipedia article, an arXiv paper or your own PDF. The browser ranks passages with BM25, and each citation in the answer opens the passage with the supporting sentence highlighted.
- 03Live
CodeLens AI
Verified Code Review
Paste code, load a file from public GitHub or paste a pull request link. A second model pass checks each review comment against the code and keeps, moves or drops it, with a reason that quotes the code.
- 04Live
VoxAI
Streaming Voice Assistant
Ask by voice or by typing. The reply streams in and the browser starts speaking at the first sentence, with the time to first audio measured and shown. Live weather (Open-Meteo) and Wikipedia lookups feed the answers.
- 05Live
DataPilot
AI Data Analyst
Ask a plain-English question about a live public dataset (USGS, Open-Meteo) or your own CSV, then follow up to refine it. The model writes a query plan and your browser runs it over every row to draw the chart.
- 06Live
ModelArena
Blind Model Arena
Send one prompt to three models, read the answers without names, then vote. A shared Elo leaderboard fills only from real votes. Latency, tokens and cost are measured, and an AI judge’s note is labelled as opinion.
- 07Live
ContentForge
Agentic Content Pipeline
Turns a topic into a cited piece, grounded in live Wikipedia and Hacker News sources, through five writing steps. Changes between Draft, Edit and Polish are shown word by word, explained, and checked against the text.
- 08Live
VisionLab
Multimodal Vision AI
Ask about a whole picture, about a box you draw on it, or about two pictures side by side. Pick an image from Wikimedia Commons or upload your own; a region question sends the crop at full resolution.
- 09Live
InsightHub
npm Download Analytics
Loads live npm download counts for one to five packages and marks the unusual days. Each spike is matched to the package’s release history, and a streamed explanation has every number, date and version checked against that evidence.
- 10Live
BrowseBot
Browser Agent with Visual Replay
Describe a web task and a model plans the steps. They run in headless Chromium on allowlisted public sites, and a filmstrip replays the real page at each step with the URL, title and text the browser read.
- 11Live
GraphScout
LangGraph Research Agent
Answers a factual question from Wikipedia with an agent loop, a draft and a critic, and lists its cited sources. You can rewind to the plan or the critic, edit it, and resume the run from that checkpoint.
- 12Live
GraphGate
GitHub Issue Triage with Approval
Triages live public GitHub issues. It looks for earlier duplicates and quotes the evidence, then proposes labels, a priority and a reply or “close as duplicate”, and waits for a human to approve, edit or reject. Nothing is posted to GitHub.
- 13Live
GraphSwarm
LangGraph Map-Reduce
Splits a Wikipedia article or pasted text into chunks, extracts points from each in parallel and writes a summary that cites its chunks. A coverage map shows which chunks the summary uses, and missed chunks are re-run once.
n8n workflows
The repository also holds five n8n workflows: lead capture with LLM scoring, support triage with vector retrieval, invoice extraction with a human approval step, data sync with a dead-letter queue and replay, and a KPI monitor with anomaly detection.