📍 NL · ⭐ 179 stars | 🔱 15 forks | 📦 48 repos

AI-Native Builder. Local-first AI agents and privacy-first products — no accounts, no tracking, open rails over rented infrastructure.

AI-Native Builder and technical leader — shipped at Meta, Vodafone, Liberty Global, and Cognizant. I take problems to working products: architecture, code, and the decisions in between.

Capability without custody. I don't build new infrastructure — I reshuffle the rails you already use (email, maps, git, chat apps, the logs your apps already write) so the leverage lands on you, not a platform. Borrowed rails never need custody of your data — privacy here isn't a promise, it's structure. No account to breach, no profile to sell, no lock-in: you keep the function and the keys, never become the product.

Always up for hard problems and good teams — let's talk.

Simple over clever. Every line must have a purpose.

Two throughlines, one principle: local-first AI agents, and privacy-first products & primitives — no accounts, no custody, no rented infrastructure.

Local-first rails for AI agents, plus tools for building with AI. My most active work.

baresuite — agent infrastructure · light enough to read in an afternoon, zero/low deps

Core — the brain, the gate, the memory

• ⚡ bareagent — Give an agent a goal, get coordinated actions — or hand it a hard one and it decomposes, fans out, verifies, and synthesizes (RLM). Replaces LangChain, CrewAI, AutoGen.
• 🚦 bareguard — One gate on every action an agent takes: allow / deny / ask-a-human — plus an opt-in harness that checks outputs against intent.
• 🕸️ litectx — Tree-sitter memory with activation decay + context-engineering verbs (write/select/compress/isolate). Ranked recall + impact over SQLite/FTS5, no LSP.

Optional reach — give the agent hands

• 🌐 barebrowse — Let agents browse the web like a human. Replaces Playwright, Selenium, Puppeteer.
• 📱 baremobile — Let agents drive Android + iOS devices. Replaces Appium, Espresso, XCUITest.
• 📨 beeperbox — Reach 50+ messengers (WhatsApp, iMessage, Signal, Telegram, Slack, …) through one MCP server — full Docker appliance (headless Beeper bundled) or a lite npx MCP layer over your own Beeper.

Building with AI

• 🤖 liteagents — Skills toolkit for AI coding assistants; learns from past sessions to work better over time.
• 💬 multis — An agent that lives in your chat apps: controls your machine, remembers conversations, searches documents. Built on the rails above.
• 🧰 agentic-toolkit — Agent workflows, automation, prompt engineering.
• 🧪 edgelms — Where does a frozen local embedder earn its keep? One model, four endpoint jobs (log/command anomaly, routing, entity resolution), measured on real data with pre-registered verdicts: it wins at narrowing/recall, loses at judgment. Zero training.
• 🦙 coding-assistant — Run small language models locally for code assistance. No cloud required.
• 🔬 relayfact — An autonomous senior-dev runner assembled from the bare suite: grounds the loop on executable verification (checks that can fail) and narrates itself as an event stream. Builds no primitives — it either graduates or gets archived. [WIP]
• 🧬 adaptlearn — Can an agent's harness — not its plan or code — be an emergent artifact that improves across runs? Assembled from the bare suite; a dumb outer loop holds the grounded close, the agent authors its own workflow and inherits what worked (verdict-gated). Sibling of relayfact — graduates or gets archived.
• 🔁 bareloop — Workflows that earn their own design, with receipts: describe a repeated job and its checkpoints, an agent authors the scaffolding, runs execute under an un-gameable gate, and the scaffolding improves across runs (verdict-gated inheritance). Graduated from the adaptlearn experiment. [WIP]

_{Lineage: aurora (archived) pioneered this — its memory became litectx, its orchestration became bareagent.}

The same local-first approach as the rails above — shipped as products people use, and as primitives to build on. No accounts, no tracking, no central trust.

Privacy products · use them today

• 📍 addypin — Turn a GPS coordinate into a short, memorable link (addypin.com/HOUSE1) or email HOUSE1@addypin.com. 12 map-app buttons, no accounts. · repo
• 📨 signedreply — Coordinate multi-party actions over email. Every reply DKIM-verified + OpenTimestamped and committed to a per-event git repo — proofs verify offline even if the service dies. · repo
• 🏛️ plato — Self-hosted forum: Reddit-shaped, 2002-operated, one program + one data file, plain-text posts. Live at ownsub.com. · repo
• 🚆 late.fyi — Email a train number → real-time platform, delay, and cancellation alerts back. Your inbox is the protocol. · repo
• 💬 ama — Ask any website anything using your existing AI subscription. Researches whole sites, translates foreign content, answers in English.
• 🔒 privpn — Your own WireGuard VPN on a VPS. Your server, your keys, no third-party provider.
• ☁️ privcloud — Self-hosted home server: photo backup, music streaming, files, remote access. One script from fresh Fedora to fully running.
• 🛡️ wearehere — All-in-one browser privacy audit: cookies, trackers, fingerprinting, dark patterns, ToS toxicity — one scan. Also ships an MCP server for AI agents.

Privacy primitives · build on them · local-first, no telemetry

• 🔑 knowless — Full-stack passwordless auth for Node.js: magic-link sign-in, no passwords, no profile stored. The auth layer under everything I ship.
• 🗂️ flightlog + pulselog — a lightweight, self-hosted server-log suite; same zero-dep JSONL dialect, read with tail/jq.
  • ✈️flightlog — in-process error capture: uncaught exceptions, rejections, and errors you hand it. The local alternative to Sentry.
  • 🩺 pulselog — external watcher: health/SSL/disk/backup checks, a weekly stats digest, rotated backups. Replaces hosted analytics + uptime monitoring.
• 🔞 8een — One-bit, unlinkable, stateless age verification — the ZK verifier the EU didn't ship. Proof in → true/false out. Replaces ID uploads and face scans. Built on google/longfellow-zk. [WIP]
• 🪪 zkagent — One passport scan → one anonymous, unforgeable tag per service: human-rooted auth for AI agents. Ban the tag and every bot behind it stays banned — no CA, no accounts, nothing stored. [WIP]
• 📶 justabit — Telecom network APIs that answer instead of disclosing: predicate + floor + nonce in → one signed, expiring bit out. Never the timestamp, country, or number behind it; the aggregator bills but can't read. Standards staging for CAMARA + AAIF. [WIP]

• 🔊 sawt — Turn any book into a multi-voice Arabic audiobook. File in, narrated audio out.
• 🧾 mailproof — Email-native coordination kernel: verify a reply (DKIM/DMARC), commit it to a tamper-evident git ledger, sequence the workflow, trigger the next email. The engine extracted from signedreply. Zero deps.

• POC first — validate the idea before committing to a design
• Simple over clever — fewer deps, less config, readable code
• Borrow and improve — every past project is a library of lessons
• Local-first, open-source only — your data stays on your machine
• Connect the dots — cognitive science, product ops, AI research — patterns transfer
• Full dev standards — the rules every project follows
• Building a JS library — the fixed ref for publishable JS libs: types (JSDoc → .d.ts → CI, no drift), the adopter context.md, doc set
• Shipping a container / MCP server — container-correctness foot-guns (signal forwarding, healthcheck-through-the-forwarder, loud-open defaults, pin+sha256) and MCP tool-surface discipline

Each project started as a question I couldn't stop thinking about — from NLP for Arabic, to map tech for developing nations, to AI frameworks that give agents memory and reasoning. I build things to understand them, and share what works.

Read the full story: doing more with less — a year of building privacy primitives and AI tooling