I build backend systems and the developer tools that make AI agents genuinely useful. Ten-plus years shipping production Python and TypeScript — I scaled the Magic Labs SDK to millions of users, built an OIDC/JWT Wallet-as-a-Service to ~$3M ARR, and helped build the CNN Magic Wall for the 2020 US election. These days I'm heads-down on agent infrastructure: MCP servers, context engineering, autonomous code-review, and the connective tissue that makes coding agents faster, cheaper, more reliable, safer, and more autonomous.
Most of what's below is a body of work I've been building deliberately rather than marketing — substance first, launch later. Based in New York City.
Backend · AI / Agent Infrastructure · MCP & Tooling · Async Python · TypeScript · Rust
The frontier work: making coding agents faster, cheaper, and less forgetful.
| Project | What it does |
|---|---|
| lessons-learned | Captures coding mistakes and injects the relevant lesson before the next tool call — with a built-in eval framework that proves which lessons actually work and auto-prunes the ones that don't. |
| mcp-exec | Reference implementation of Anthropic's Code execution with MCP (Nov 2025). Reuses Claude Code's own sandbox and config — so isolation comes free — and keeps intermediate tool output out of the context window entirely, on every workflow, not just multi-tool ones. Up to 99% fewer tokens. |
| ai-review-bot | Autonomous multi-agent PR reviewer. Waits for every other review bot to finish and indexes their findings to avoid duplication, then dispatches 5+ specialized sub-agents (scaled to the PR) to review and suggest. The full review → re-review → approval loop runs end-to-end, with QA gates and a feedback loop that folds in reviewer remarks and self-improves over time. |
| ctx-tree | Tree/graph-based persistent context store for long-running agent sessions. Pairs with mcp-exec to keep raw output out of the window — returns identifiers, snippets, and hints instead of full payloads, so the agent pulls exactly what it needs and works far longer before context fills. |
| idiomatic | Language-agnostic idiom-enforcement gate for AI agents and humans — Rust core, sub-100ms autofix, Python + Node bindings. |
| ai-listings | AI-powered product-listing generation tool. |
| agent-skills | A large, growing library of agent skills for Claude Code & Codex — spanning Git/GitHub/Actions, Vercel, Serena, Dolt, DDD & SOLID, cross-platform plugin development, docs tooling, and agentic best-practices. Every skill auto-updates weekly from upstream docs to stay current. |
| personal-skills-showcase | A public window into how I build — my "Phosphor" design system and engineering standards, curated from my private skills library as Claude Code / Codex agent skills. |
| cc-recall | Discoverability + enrichment for Claude Code session transcripts — fixes a Claude Code bug that leaves ~90% of sessions undiscoverable, so past work can actually be found and reused. |
| agent-marketplace | My curated marketplace & storefront — where I publish, version, and showcase every plugin, skill, and tool I ship for Claude Code & Codex CLI. Also home to the idea-nursery, where nascent macOS tools & automations incubate before earning their own repo. |
| anti-compact | Intercepts automatic compaction and halts it, generating a structured session handoff instead of lossy compression. |
| edgelite | EdgeDB-style developer experience on PGlite — SDL schema, typed TS query builder, migration CLI. |
| cc-vscode-ext | A clean, documented reimplementation of the Claude Code VS Code extension — reverse-engineered, with a number of improvements over the original. |
| pieces-dev | pnpm monorepo of PiecesOS developer tools — a monitoring platform (daemon + SDK + CLI + editor extensions), an LTM context injector for VS Code, and gap-reconstruct. |
| agent-improvement | Centralized hub for Claude/agent workflow improvements — improvement plans (AIPs), hooks, and tooling. |
Async-first Python and framework-level tooling. FastAPI + SQLModel for CRUD; DDD/CQRS for the complex stuff.
- quart-sqlalchemy ⭐ — Production async SQLAlchemy integration for Quart.
- python-banshee — My port of Banshee: a command dispatcher & message bus (Commands / Queries / Events / Handlers) for CQRS-style Python services.
- pricing — Locale-aware, CLDR-backed pricing, currency formatting & exchange.
- quart-depends · python-quart-injector — FastAPI-style dependency injection for Quart.
- redis-bloom-ttl — Rust Redis module: a bloom filter with per-entry TTL.
- hashidtools — Async hash-ID generation components.
Former Web3 technical co-founder; built wallet infrastructure and the primitives underneath it.
- coinaddr ⭐ — Cryptocurrency address inspection & validation for Python.
- btcpay-python ⭐ — The first & canonical Python SDK for BTCPay Server, adopted into the official BTCPayServer org.
- base58check ⭐ — Base58Check encoding/decoding.
- docker-geth — Dockerized Geth + Swarm (Ethereum) node wrappers.
- spinup-ts · spinup-py — Opinionated, very strict project scaffolds that bake my agentic-development best practices into each language (extensive linters + code-quality gates). Using one forces the AI to write better, more maintainable code that needs less review (Biome/Vitest/Husky · uv/ruff/wps).
- safaribooks — Revived a defunct fork: rebuilt on O'Reilly's current APIs + cookie auth, made it async, and published it to PyPI (
safaribookshelf). Beyond EPUB, it now exports agent-ready Markdown built to be consumed by AI agents or dropped straight into skills. - resume-builder ⭐ — Dynamic LaTeX résumé pipeline from a JSON Resume schema.
- jobsearch-tracker — Personal job-search pipeline tracker.
- ext-tool — Manage VS Code / Cursor extensions from one source-of-truth TOML registry.
- gstack — Forked the original and wrapped it as a Claude Code plugin, making the 23-tool setup far easier to install and use.
- tmpld — Jinja2 templating on steroids for Docker containers.
- dcgoss —
dgossvariant driven by docker-compose files. - testdocker — Unit testing for Docker containers.
- devcontainer — Reusable dev-container configurations.
- gpgkeyring — Programmatic GPG keyring management.
- cookiecutter-ds — Cookiecutter template for Python data-science projects.
☸️ Cloud-native & Kubernetes — deeper systems catalog (2016–2020)
- kubed — An opinionated, human-friendly Kubernetes client.
- pycaps ⭐ — Linux capabilities for Python.
- overridehostname ⭐ —
gethostname(2)interposer. - docker-couchdb ⭐ — Dockerized CouchDB 2.0.
- couchdiscover — CouchDB 2.0 autodiscovery on Kubernetes.
- rpi-cluster-bootstrap — Bootstrap for a Raspberry Pi k3s cluster.
- kubewait — Init container that pauses startup until dependencies are ready.
- pyrkube — Read-only
pykubewrapper tuned for a different focus. - kube-ssl — Kubernetes TLS helpers.
- k8s-kernel — Kubernetes-optimized Linux kernel.
- container-utils — Shell utilities for containers.
- debian — Minimal Debian for containers with Kubernetes extras.
🔩 Hardware, home automation & earlier work
- homepod-update — Turns a jailbroken 1st-gen HomePod into a bidirectional voice satellite — swap Siri for Claude / GPT / any LLM via Home Assistant or an OpenClaw cluster. A
mediaserverdtweak streams beam-formed mic audio (Wyoming protocol), with a Go management dashboard — enabled by a checkm8 jailbreak + a custom pogo-pin USB dock adapter. - luxafor — Python driver + internet-health monitor for the Luxafor Flag USB HID device.
- brca-crawler — Scrapy-based crawler that harvests county property-parcel records into structured data.
- video-pro — Application for managing video content at scale.
- lifx-demo — LIFX Wi-Fi LED control demo (Ruby).
Open to senior backend & AI-infrastructure roles, custom AI development & consulting engagements — and collaborators at the frontier.
📫 me@joeblack.nyc · 🌐 joeblack.nyc · 𝕏 @joeblackwaslike