Open Desk is a four-agent team built on ClawMax. You give it a public GitHub repository. Scout reads the project and describes it in plain language. Pathfinder finds the contributing rules. Suggester writes three ways to help, each tied to a real issue or file. Evaluator checks that the ideas match the repo and those rules, then stops and asks before anything is posted.
The workflow is Map the Repo, Find the Path, Suggest Contributions, then Check and Ask. In the demo, the crew reviewed Maximilien-ai/clawmax and suggested work on issues #225, #223, and #162. The evaluator passed all three. Nothing was posted upstream.
ScamShield is a multi-agent AI assistant built with ClawMax to help everyday people identify phishing, scam messages, and online fraud without needing technical knowledge.
I created a custom ClawMax team with four specialized agents and four workflows:
1. **Scam Analysis** - Detects suspicious messages, phishing indicators, impersonation, and social engineering.
2. **Independent Verification** - Checks suspicious claims and separates verified evidence from assumptions.
3. **Safety Review** - Evaluates findings and assigns a risk level and confidence level.
4. **Final User Guidance** - Explains the results in simple language and recommends safe next steps.
The demo uses a fictional PayPal phishing SMS and shows how the four workflows produce a user-friendly safety report.
Alongside ScamShield, I contributed improvements directly to ClawMax during the hackathon at Qualcomm.
I reported 24 GitHub issues and submitted 23 pull requests, including 8 security fixes addressing unsafe code execution, file path vulnerabilities, API key exposure, and other security risks.
Additional contributions improve workflow execution, reliability, accessibility, and dashboard usability.
All 23 pull requests passed GitHub Actions CI and are available for maintainer review and merge.
The goal is to make AI agents both more useful for non-technical people and safer to operate.
# Doom Cart
See what your shopping costs the planet. Doom Cart is a one-screen dashboard that shows how everyday purchases affect climate, water and mineral reserves, which online stores emit the most, and which swaps actually help.
Built for Patch the Conference (Bright Data, San Francisco, 2026-10-24).
## What it does
- **Live counters:** global purchases of 7 everyday products, ticking per second, plus days since Earth Overshoot Day.
- **Globe:** where a product's materials come from, with arcs to a US buyer.
- **Search:** type a product or paste an Amazon, Walmart, Target, Shein or Temu link to see its footprint.
- **Selling now:** real Amazon and Walmart listings with "bought in past month" badges and estimated CO₂ per month.
- **Stores:** reported emissions for Walmart, Amazon, Target, Shein, eBay and Temu.
- **Action forecast:** tick swaps (refurbished phone, plant-based burger, second-hand clothes) and see your CO₂ to 2030, then scaled to a million people.
- **Resource clocks:** when known cobalt, nickel, copper and lithium reserves run out at a chosen demand growth.
- **Receipt:** your year of shopping, against a 2.5 t yearly climate budget.
## Data
| Data | Source | File |
|---|---|---|
| Per-item CO₂, water, cobalt | Apple product reports, ADEME, Poore & Nemecek 2018, Water Footprint Network, MIT | `data/footprints.json` |
| Global units per year | IDC, Omdia, UNEP, USDA, World Footwear Yearbook | `data/footprints.json` |
| Store emissions | Each company's 2025 sustainability report | `data/retailers.json` |
| Mineral reserves | USGS Mineral Commodity Summaries 2026 | `data/footprints.json` |
| Live listings | Amazon and Walmart US search pages, scraped with Bright Data | `data/live.json` |
| Overshoot date | Global Footprint Network, 2026 | in page |
Known weak spots: the burger uses the global beef average (US beef is lower), coffee CO₂ and water assume different doses per cup, sneakers and laptops have no water figure, and cobalt per device comes from 2016 reporting. Store totals are not like-for-like because each company counts its supply chain differently.
SkyWatch shows every aircraft in the air right now on a live world map, and lets you ask an AI agent what any of them actually is.
A dot on a flight map tells you where a plane is, not what it is. Click any plane and hit "Brief me": Claude researches it on the live web and returns the airline, route, aircraft type and status, with sources. You can also ask questions about the traffic on screen ("What's the fastest plane and where is it going?").
HOW IT USES CLAUDE + BRIGHT DATA
- Claude is the brain: a multi-step tool-use agent that decides what to search and which pages to read, runs independent calls in parallel, and says plainly when it can't confirm something. Every step streams live to the UI.
- Bright Data is the eyes: the SERP API gives Claude live Google search (web_search tool); Web Unlocker reads flight-tracker and airline pages that block bots, need JavaScript or show CAPTCHAs (fetch_page tool). It also fetches the worldwide ADS-B feed from OpenSky.
FEATURES
- Live world map: 10,000+ aircraft, colored by altitude; dots when zoomed out, rotated plane icons when zoomed in
- Continuous motion: planes are dead-reckoned from real speed and heading and glide between updates
- Brief me: cited flight briefings from the live web
- Ask the skies: chat grounded in the on-screen traffic snapshot
BUILT FOR SPEED
- One canvas renders 10k+ planes smoothly
- preload.py caches a world snapshot, so the server starts instantly
- Live data refreshes every 30 s in a background thread; panning and zooming are served from memory
Built tonight with Claude (including Claude in Chrome to set up Bright Data zones and GitHub).
Incident Lens turns a plain-text research question into an auditable public-evidence packet for injury investigations. Claude parses the question into typed inputs; Bright Data SERP + Web Unlocker (plus CPSC and SF311 open APIs) discover and retrieve official sources, each stored with hash and transport log; Claude extracts facts with verbatim quotes; our own code runs the checks — model match, serial-range, publication date vs incident date, distance, before/after. Output: findings, conflicts pre
PythonAnthropic SDK (Claude Sonnet 4.5 structured outputs)PydanticBright Data SERP + Web UnlockerCPSC Recall APISF311/Socrata SoQLCensus GeocoderLeaflet/OSM
