Portfolio / active systems
A working catalogue of agent infrastructure, local-first command-line tools, product experiments, libraries, and editors. The newest work forms one connected evidence stack—from repository creation to browser and terminal verification.
Browse every repositoryCurrent system
Five independent tools that connect planning, context, and behavioral evidence.
Project directory ↓A local-first control plane for software-engineering agents. Cortex preserves goals, evidence, bounded changes, verification results, and durable memory so long tool-using sessions stay inspectable and recoverable.
17-tool compact MCP profile · 24-tool operator profile
Turns a small bob.yaml product contract into a reviewable repository plan, then applies only the files it can prove it owns. Bob is model-free, drift-aware, and designed to make generated foundations safe to review and maintain.
Atomic ownership checks · 6 typed MCP tools
Define MCP servers once, expose their tools through a single namespaced gateway, and safely synchronize the same configuration into every supported coding-agent harness. Local usage intelligence makes the invisible tool layer observable.
12 agent harnesses · 7-tool lazy management surface
Local-first behavioral browser-spec layer for coding agents. Specs define intent + outcomes as the behavior contract and steps as repairable hints. The same spec can run from the CLI, through the MCP server, or be exported to Playwright. Captures DOM snapshots, screenshots, console, network, and outcome evidence into one agent-readable artifact pack.
Browser evidence packs · locator-drift healing
Local-first behavior runner for terminal applications. Launches a target command inside a real PTY, drives it from YAML/JSON steps, evaluates outcomes against a deterministic virtual terminal screen, and writes self-contained artifact packs. Full xterm control set support including SGR colors, OSC 8 hyperlinks, and mouse input. Runs on Unix PTYs and Windows ConPTY.
Unix PTY + Windows ConPTY · JUnit and BATS interchange
01 / Command line
Evidence-guided agent kernel for long-running software work
02 / Command line
Deterministic repository factory for agent-native tools
03 / Command line
One local gateway for every MCP server and agent harness
04 / Command line
Browser behavior spec runner for coding agents
05 / Command line
Terminal app behavior runner with PTY and artifact packs
36 shown
A local-first control plane for software-engineering agents. Cortex preserves goals, evidence, bounded changes, verification results, and durable memory so long tool-using sessions stay inspectable and recoverable.
17-tool compact MCP profile · 24-tool operator profile
Turns a small bob.yaml product contract into a reviewable repository plan, then applies only the files it can prove it owns. Bob is model-free, drift-aware, and designed to make generated foundations safe to review and maintain.
Atomic ownership checks · 6 typed MCP tools
Define MCP servers once, expose their tools through a single namespaced gateway, and safely synchronize the same configuration into every supported coding-agent harness. Local usage intelligence makes the invisible tool layer observable.
12 agent harnesses · 7-tool lazy management surface
Paste any URL and LinkGlow classifies every link on the page — internal or external, follow, nofollow, sponsored, or UGC — and renders an interactive glow view highlighting each link over the real page. The free analyzer is the front door; the product is placement monitoring: track the backlinks you've earned or paid for and get alerted when one breaks, disappears, or silently flips to nofollow.
Write API tests as plain .http files with 26 assertion operators, built-in stress testing, and a web UI. Single binary, CI/CD-ready.
Write Vue 3 components that render real native views — no WebView, no compromise. Supports Composition API, 20+ built-in components, native modules, hot reload, and cross-platform from a single codebase.
A declarative programming language that compiles .bp files to runnable TypeScript projects. Write a spec, get a working API powered by Hono, Drizzle, and Zod. No boilerplate, no lock-in.
Get free AI tarot readings with 6 spread types — Single Card, Yes/No, Three Card, Love, Career, and Celtic Cross. Features custom pixel-art Major Arcana illustrations and streaming interpretations powered by Claude AI.
Fast CLI for capturing and organizing notes with tagging, full-text & semantic search, and a memory system for AI agents via MCP.
Note-taking plugin with wiki-links, daily notes, quick capture, and Telescope integration. Notes live in a centralized vault accessible from any project.
Inline ghost-text code completion powered by Claude Haiku. Progressive acceptance, LSP/Treesitter context awareness, and minimal latency.
Combines semantic code search with local LLM intelligence via Ollama. Ask questions about your codebase and get contextual answers.
Embeddable vector database for Go. Single-file persistence, HNSW indexing, hybrid search (vector + BM25), streaming results, metadata filtering, and 55 MCP tools. Designed to be embedded in other Go applications with zero external runtime dependencies.
Local-first semantic code search powered by embeddings. Index your codebase and search by meaning, not just keywords. Defaults to local Ollama embeddings with optional cloud providers. Hybrid mode, MCP server, studio UI, and similar code finder.
Save, restore, compress, index, search, and connect agent-generated files and folders back to code. A single local binary keeps bulky evidence addressable without turning repositories into artifact dumps.
14 MCP tools · automatic compression above 10 MB
Dead-simple local secrets manager for developers and AI agents. Single Go binary with AES-256-GCM encryption, Argon2id key derivation, versioned secrets with history and rollback, X25519 recipient sharing, .env.encrypted files safe to commit, git-filter transparent interpolation, local unix agent for passphrase-free daily use, relational search, interactive terminal studio, access policy YAML, and 49 MCP tools.
49 MCP tools · six OS/architecture release targets
Local-first Go CLI that takes a screen recording of a bug and produces frames, OCR text, transcripts, metadata, and a timeline that connects what was visible with what was said. Designed for QA engineers, developers, and coding agents that cannot watch a video directly but can inspect files and JSON.
Local-first behavioral browser-spec layer for coding agents. Specs define intent + outcomes as the behavior contract and steps as repairable hints. The same spec can run from the CLI, through the MCP server, or be exported to Playwright. Captures DOM snapshots, screenshots, console, network, and outcome evidence into one agent-readable artifact pack.
Browser evidence packs · locator-drift healing
Local-first behavior runner for terminal applications. Launches a target command inside a real PTY, drives it from YAML/JSON steps, evaluates outcomes against a deterministic virtual terminal screen, and writes self-contained artifact packs. Full xterm control set support including SGR colors, OSC 8 hyperlinks, and mouse input. Runs on Unix PTYs and Windows ConPTY.
Unix PTY + Windows ConPTY · JUnit and BATS interchange
Text-based multiplayer online battle arena where strategy matters more than reflexes. 5v5 turn-based combat on 4-second ticks with 18 programming-themed heroes, 40+ items, talent trees, draft phase, deny system, buyback, vision game, and anti-cheat with leaver detection and low-priority queue.
Upload a two-hour interview and ask for 'the three best questions the host asked' or 'every time they mention pricing' — ClipIt returns exact time ranges with a written rationale for each match, backed by the transcript. A retrieval tool, not a virality machine. Shared clips carry their evidence — transcript, rationale, and timecode — and video ingest, transcoding, and streaming run on Mux.
Helps musicians structure their practice with interactive tablature, audio playback, and progress tracking. Built with Nuxt 4 and powered by AlphaTab for high-fidelity tab rendering and Tone.js for audio synthesis.
Document reservation platform for managing document workflows with role-based access and real-time status tracking. Built with Bun, Turborepo, TypeScript, and PostgreSQL.
Teaches programming through fill-in-the-blank exercises over real code — the middle ground between multiple choice and a blank editor. Every submission is graded by actually executing the exercise's test suite inside an isolated sandbox microVM, and an SM-2 spaced-repetition engine schedules reviews. Tracks cover TypeScript, Python, Go, Rust, Vue, and React, including exercises for working effectively with AI coding agents.
Tactical card game that combines tarot symbolism with trading card game mechanics. Built with Next.js 15 and React 19.
Git visualization tool designed for both humans and LLMs. Clean, parseable output of repository state and history.
Lightweight native macOS code editor built with Swift and SwiftUI/AppKit. VS Code-like editing experience with multi-tab editing, syntax highlighting for 20+ languages, LSP integration with semantic-token highlighting, DAP debugger, a real PTY-backed interactive terminal via SwiftTerm, project-wide search, command palette, git integration with push/pull/fetch and ahead-behind sync, crash recovery with incremental editor indexing, code folding, minimap, and session persistence. Multiple themes including Nord, GitHub Light, and Dracula.
A terminal coding agent with approval-gated tools, MCP integrations, durable goals, expert consultation, and optional workspace-scoped memory. It runs against local Ollama models and keeps resumable sessions in SQLite.
NORMAL, PLAN, and AUTO modes · lossless SQLite resume
Modern terminal code editor with a VS Code-like experience. Multi-tab editing, syntax highlighting for 40+ languages via Chroma, LSP support, file tree sidebar, git panel, text and semantic search, Lua plugins, live file watching, immutable rope-backed undo/redo, full mouse support, and Nord theme with 30+ styles.
The same real-time runtime data through a polished terminal UI, stable JSON commands, and an MCP server. Monitor covers CPU, memory, disks, network, sensors, services, containers, and processes on macOS and Linux.
9 Studio tabs · 8 confirmation-aware MCP tools
Terminal UI for managing macOS audio devices built with Bubble Tea and CoreAudio. Switch defaults, adjust volume, toggle mute — all from your terminal. Real-time event-driven updates, hot-plug support, per-channel volume fallback for USB interfaces, and Catppuccin Mocha theme.
AI-powered book writing platform built with Nuxt 3. Context-aware writing assistant powered by Claude API with two model tiers, collaborative character and world-building tools, TipTap rich text editor, manuscript management with export, AI streaming via H3 EventStream, Drizzle ORM with SQLite, and Zod validation.
Library operator and stack readiness CLI and MCP server for the shared ~/.agents tree, built with Go and Cobra. Manages skills and agent profiles on disk (the same layout local-agent loads), seeds profiles from templates, and probes companion tools (bob, cortex, mcphub, codemap, vecgrep) using real binary names with tiered health checks. Ranked suggestion engine, append-safe analytics log, deep readiness doctors, and a full MCP stdio server.
Morse code learning app with a native iOS app in Swift 6 and SwiftUI, a Bun/Hono/GraphQL API backend, and a Nuxt companion web app. Shared domain logic in TypeScript and Swift packages kept in parity. Features SRS, gamification with XP/streaks/hearts/unlocks, 15+ achievements, challenge modes, curated learning paths, Core Haptics, and CoreAudio Morse playback.
Turns a subset of a Docker Compose project into a disposable environment on DigitalOcean. Compute and data live in separate Pulumi stacks, so environments can be destroyed and recreated freely while persistent volumes survive. Services are never exposed to the public internet — only SSH is reachable, through an explicit firewall — and every run shows a redacted preview of what it will provision before touching the cloud. Distributed via npm and a one-line installer, with a hosted control plane for teams.
Disposable compute · protected data stacks
Measures and improves how a brand appears in AI-powered search surfaces such as ChatGPT and Perplexity. Every observation is tied to the exact model that produced it, with query coverage, citations, confidence, campaign planning, and page recommendations that stay traceable back to evidence. Multi-tenant SaaS with row-level security on every table, separate database roles for runtime, billing, and migrations, durable workflow runs with audited recovery paths, and a deterministic public demo.