Agent infrastructure engineer. I build the layer between you and the agent loop.
Columbus, Ohio · edwardkubiak.com · LinkedIn · edward.kubiak.dev@gmail.com
Open to roles on agent infrastructure, developer tools, and Claude Code platform teams.
Lately I’ve been focused on the operational side of multi-agent systems — routing, memory, observability, policy gates — the kinds of details that make the difference between an impressive demo and something reliable enough to leave running on its own.
For the last six months, I’ve been building CAST, a local-first OS layer for Claude Code, plus a handful of smaller supporting packages around it.
claude-agent-team · Bash · MIT · v7.0 — backend lockdown
Claude Code ships with a powerful agent primitive and almost no scaffolding around it. CAST is the scaffolding: 22 specialist agents, hook-driven dispatch, model-tier routing, per-agent persistent memory, and a 27-table SQLite event store that makes the whole loop legible.
brew tap ek33450505/cast && brew install castA few of the design decisions worth defending:
- Hook-driven dispatch over polling. Claude Code emits lifecycle events (
SessionStart,PreToolUse,SubagentStop,Stop). CAST hangs all routing, telemetry, and policy off those — no daemon, no background loop, no missed events. The harness is the orchestrator. - Local-first, SQLite everywhere. Agent runs, routing decisions, memory, quality gates, hook output — all in
~/.claude/cast.db. No cloud round-trip. The dashboard is just a SELECT away. Data is yours; observability is yours. - Per-agent memory, not shared context. Each agent keeps its own scoped memory under
~/.claude/agent-memory-local/<agent>/. Cellar Door (below) extends this to a typed shared store when agents do need to coordinate — but the default is isolation, because shared context bleeds. - Policy gates that hard-block. Branch-protection bypass attempts, force-pushes to main, raw
git commit— the hook layer refuses them. Quality is enforced at the seam, not by hoping the agent behaved.
990 BATS tests across 111 files cover the shell surface. The framework treats its own correctness as a first-class concern.
Each package below is independently installable and stands on its own. CAST is what happens when you wire them together.
| Package | One line |
|---|---|
cast-desktop |
Tauri 2 desktop app — every CAST signal in one window. 11 dashboard views, native PTY terminal (xterm.js + Rust), Cmd+K palette, 6 themes. Local-first against ~/.claude/cast.db. v0.1.0. |
claude-code-dashboard |
React 19 + Express + SQLite observability UI for the agent loop. Sessions, agents, hook health, memory browser, SQLite explorer. v2.2.0. |
cellar-door |
Typed shared memory for local AI agents — model-agnostic, Claude + Ollama. v0.3.0. |
cast-hooks |
Standalone hook scripts — observability, safety, dispatch. Drop into any Claude Code setup. v0.2.0. |
cast-memory |
Persistent agent memory with FTS5 search, relevance scoring, MCP server. v0.3.0. |
cast-dash |
Terminal UI dashboard — Python + Textual. htop for CAST. v0.2.0. |
cast-time |
Gives Claude Code a clock — local time + timezone injected at session start. v0.1.0. |
cast-routines |
Schedule autonomous Claude Code routines — daily briefings, inbox triage, release celebration — via YAML + cron. v0.1.0. |
cast-doctor |
Read-only health check for any Claude Code install — validates hooks, MCP, agent frontmatter, cast.db schema. v0.1.0. |
cast-claudes_journal |
Cross-session continuity for Claude — Obsidian vault, wiki-link graph, SessionStart context injection. v0.2.0. |
cast-observe |
Session cost + token-spend tracking. No framework required. v0.2.0. |
cast-security |
Policy gates, PII redaction, audit trail, cast-audit consolidation. No framework required. v0.3.0. |
cast-parallel |
Plan execution split across parallel git worktrees. v0.1.0. |
All open source. All Homebrew-installable. All written to be useful without the rest.
Counts and versions on this page refresh dynamically from ~/Projects/personal/ via scripts/refresh-stats.sh.
Bash · Python · TypeScript · React 19 · Express 5 · SQLite · BATS · Vitest. macOS first, Linux supported. Anthropic API + Claude Code Agent SDK.
Currently writing about agent observability, memory architecture, and what production multi-agent systems actually need. Reach out: edward.kubiak.dev@gmail.com.