Florian Narr

A Vercel Labs template that runs a marketing team as an eve agent hierarchy — one routing lead, five specialists that never delegate further, and an approval layer built from allow-lists and substring matching instead of a permissions framework.

An open-source, agent-first CRM where the research agent runs on its own schedule, works a leased task queue instead of a cron loop, and refuses to write a fact to a contact record unless the evidence clears a floor.

Vercel Labs' toolkit for building desktop apps in TypeScript and a custom markup language that compiles down to a Zig binary — no Electron, no WebView, no JS runtime shipped.

pdfme generates PDFs from JSON templates using a plugin-per-field-type architecture on top of a forked pdf-lib.

Hyvor Relay is a self-hosted alternative to SES and Mailgun — Symfony backend, a Go worker fleet that speaks raw SMTP, and Postgres doing double duty as both the queue and the DNS record store.

Caveman started as a skill that made agents talk in terse caveman-speak.

Ponytail injects a 'lazy senior dev' ladder into every prompt and subagent so Claude Code, Codex, and 18 other agent hosts stop reaching for a library when a one-liner does.

A German-language Claude plugin marketplace covering 235 legal practice areas, built with regression tests against past hallucinations and a rubric-graded eval harness borrowed from Harvey's LAB benchmark.

A Python/TypeScript library from Braintrust that turns LLM-as-a-judge scoring into YAML files, matches lists with the Hungarian algorithm, and now has to paper over the GPT-5 Responses API.

ZeroFS serves an S3 bucket as a POSIX filesystem over NFS and 9P, and as a raw block device over NBD — no database, no metadata server, just a log-structured store built on a forked SlateDB.

PageIndex replaces the embeddings-plus-vector-DB stack with a tree index an LLM navigates by reasoning, and backs every merge/expand decision with an explicit, cross-checked search-cost formula.

The Pierre Computer Company open sourced the diff/code editor and file tree components behind its product — a piece-tree text buffer with cache-aware tokenization and a struct-of-arrays file tree built for million-node repos.

A Claude Code skill that catalogs the phrases and sentence structures that give away AI-written prose, then scores drafts against them before they ship.

Coral is a local-first Rust runtime that maps YAML-declared API specs to SQL tables, then rewrites cross-source JOINs into deduplicated, parameterized API calls through a custom DataFusion optimizer.

A Zed extension that ships syntax highlighting for HelixDB's HelixQL query language, built around a local check script that catches broken tree-sitter queries before Zed's own extension CI does.

A Claude Code plugin marketplace bundling 204 agents, 244 skills, and 204 commands for GTM teams — the marketplace validator is more interesting than most of the prompts it checks, and the docs contradict their own numbers.

Zenbu.js ships Electron apps as raw, uncompiled source that users can fork and hot-patch on their own machine, using a Babel transform that turns every top-level function into a hot-swappable, plugin-interceptable reference.

files-sdk gives S3, GCS, Azure, Vercel Blob, Dropbox, and 40+ other storage backends the same upload/download/list surface, with capability flags instead of silent no-ops when a provider can't do something.

TablePro is a free, open-source macOS database client written in SwiftUI that connects to 18+ databases through a plugin architecture — no Electron, no JDBC, sub-1-second startup.

HelixDB is an open-source graph-vector database written from scratch in Rust that lets you define typed schemas and compiled query endpoints — no separate vector store, graph DB, or application layer needed.

deepsec runs regex matchers to find candidate files, then unleashes Claude or Codex to investigate them — a two-stage pipeline that separates cheap pattern matching from expensive AI reasoning.

Sandcastle is a TypeScript library that runs Claude Code (and other AI agents) inside Docker, Podman, or Vercel sandboxes, giving each agent its own git worktree and merging commits back cleanly.

An open-source agent that reads papers, writes training code, and launches GPU jobs on HF infrastructure — all from a single CLI command.

A repo of 30 Go coding challenges that ships its own web UI and test execution engine — fork it, spin up localhost:8080, and get automated test feedback without touching any CI system.

codemap gives LLMs instant architectural context about your codebase — hub files, dependency graphs, diff analysis — without burning your entire context window on file dumps.

A Rust CLI that validates and executes GitHub Actions workflows on your machine, with a full expression evaluator, four runtime modes, and a TUI — so you stop discovering broken CI after the push.

A CLI that runs 230+ SEO, accessibility, performance, and security checks against any URL and emits LLM-optimized output designed to be piped directly into an AI coding agent.

A CLI from Vercel Labs that reads your project's package.json and config files, detects which frameworks and tools you're using, then fetches matching agent skills from skills.sh.

Open Agents is an open-source template from Vercel Labs that wires together a durable workflow engine, isolated sandbox VMs, and a multi-model gateway into a full coding agent system you can fork and run on Vercel.

spawn-agent wraps locally-installed coding agents — Claude Code, Codex, Cursor, Copilot — as Vercel AI SDK providers over the Agent Client Protocol, so you can drive them programmatically with the same API you use for hosted LLMs.

wpklx discovers your WordPress site's REST API schema at runtime and generates CLI commands automatically — no hardcoded endpoints, no manual updates when plugins add new routes.

dfseo wraps the entire DataForSEO v3 API in a single Go binary with on-disk caching, TOON output, and JSONPath filtering — designed specifically so LLM agents get small, deterministic payloads instead of multi-thousand-token envelope soup.

Agent Vault is an open-source HTTP credential proxy from Infisical that intercepts your agent's outbound API calls and injects secrets at the network layer — so the agent itself never sees them.

wterm is a web terminal emulator that renders to the DOM using a Zig-compiled WASM core for near-native VT parsing performance, with a React component, WebSocket transport, and dirty-row rendering baked in.

A reference platform for cloud coding agents, Peter Steinberger on AI slop without taste, a terminal for React apps, a solo-founder time audit, launch directories, and a Claude Managed Agent on Vercel.

A fast Go CLI that wraps the Google Search Console API — JSON output, local disk cache, quota tracking, and a built-in agent skill for driving it from LLMs.

Magika replaces heuristic-based file type detection with a tiny deep learning model, achieving ~99% accuracy across 200+ content types in milliseconds on a single CPU — and it's already running at Google scale.

Toprank is a Claude Code plugin that wires Google Search Console and Google Ads into your agent — letting Claude read real traffic data, identify waste, and make account changes directly.

NotFair is an open-source Claude Code / Codex plugin that splits SEO, GEO, and paid-ads work into 45 SKILL.md procedures, each with its own LLM-judge eval suite instead of one giant marketing prompt.

RTK is a single Rust binary that sits between AI coding agents and the shell, filtering and compressing command output before it hits the LLM context — turning 2,000-token git status calls into 200-token summaries without the agent noticing the rewrite.

remocn is a shadcn-style registry of copy-paste Remotion components.

Google Research's pretrained foundation model treats time series forecasting like language modeling — a decoder-only transformer with patched input tokenization, reversible instance normalization, and autoregressive decoding that handles up to 16k context…

An opinionated starter kit that layers priority maps, auto-resolution policies, and cron-driven skills on top of OpenClaw to build a persistent AI chief of staff.

A pure-arithmetic text layout engine that side-steps DOM reflow entirely, handling CJK, bidi, emoji, and soft hyphens across every browser.

A meta-prompting and context engineering system that prevents quality degradation as Claude Code fills its context window, using spec-driven phases, subagent orchestration, and a four-type gate taxonomy.

A desktop app that runs multiple coding agents in parallel, each isolated in its own git worktree.

A SwiftUI app that scans your filesystem for AI coding agent skills across Claude Code, Cursor, Codex, Windsurf, Amp, and more — then lets you browse, edit, and organize them in one place.

A collection of 35 markdown skill files that teach AI coding agents how to do conversion optimization, copywriting, SEO, and growth engineering.

A SwiftUI app that turns Claude Code into a multi-agent workflow — cards flow from backlog to done while sessions, worktrees, tmux terminals, and PRs stay linked automatically.

OpenUI ships its own streaming-first language for model-generated UI — a hand-rolled lexer, parser, and reactive runtime that cuts token usage in half compared to JSON-based approaches.