Projects
Things I've shipped: a book, and open-source tools (all MIT). Each one came from a real problem I hit building Forge or working with Claude Code daily.
-
The Agentic Engineer Book
My first book: twelve durable principles for building production systems with LLMs in the loop.
Each principle is anchored to a real incident from something I shipped. A durable body meant to outlast the next few model generations, plus a free, versioned appendix of 40 war stories on GitHub. Out now on Leanpub, pay what feels right.
-
framesmith Active
An open-source MCP server that gives AI coding agents a visual design canvas.
Scene graph → HTML/CSS → Puppeteer screenshots. Your agent composes UI, you review in a Figma-style browser viewer, you align on the picture before any framework code gets written. v2.4.0 on npm, with phone screens as a first-class canvas type. Works with Claude Code, Codex, Cursor, Windsurf, and any MCP client.
-
Reeve Active
A Ruby gem that adds per-record authorization and a tamper-evident audit ledger to Rails MCP tools.
Reuses your existing Pundit policies, fails closed, rolls back denied writes, and records every guarded call in an HMAC-chained ledger. Plugs into any MCP server, custom ones included, through a single Reeve.invoke call. v0.8.0 on RubyGems, running in production on a MagmaLabs CRM.
-
skill-atlas Active
A Claude Desktop mod that shows every skill and subagent you have, in a pane next to the chat.
Two views: what the current session can use (project, global and plugin skills, with project ones winning on name clashes), and a map of where every skill lives across all your repos. Ported from coide in a day. Read-only, installs with a symlink, v0.5.2.
-
banneret Active
A portable sub-agent pipeline kit: read-only scouts, one builder, and verifiers that try to prove the work wrong.
Canonical scout, verifier and cold-reader contracts plus a staged pipeline playbook, compiled to Claude Code (agents and a /pipeline command) or Codex (TOML agents and a managed AGENTS.md block). Every rule traces to a real trial run. Zero dependencies, v0.3.0 on npm.
-
verderer Active
A rule ledger for AI agent harnesses: per-rule provenance for CLAUDE.md and AGENTS.md.
Inventories every harness rule with its date and origin (git blame when the file is tracked, its own history when it is ignored), records dead / derivable / needed verdicts from ablation runs, and flags them stale when the repo moves to a new model. Never deletes a rule. Zero dependencies, v0.2.0 on npm.
-
claude-code-cookbook Active
Curated recipes for Claude Code: hooks, slash commands, subagents, settings.
Each recipe is a folder with a README and copy-pasteable config — built from real workflows I actually use.
-
claude-orchestration-pack Active
A multi-agent pack for Claude Code that runs without a supervisor sub-agent.
Architectural take: peer agents coordinating via shared artifacts, not a top-down orchestrator. Cheaper, more debuggable, harder to design.
-
coide Active
An Electron/React desktop GUI for Claude Code.
My first Claude Code project. Built to scratch an itch around session management and visibility.