Work
What I build and do. One paid product (Armada, in the Atlassian Marketplace). The rest open source on GitHub. Most are Go.
PAID / 01
OPEN SOURCE / 13
Nomi
active · Go + Tauri
Local-first AI coding agent. Plan review before execution, capability-gated tools, BYO LLM (Ollama / Anthropic / OpenAI). Tauri desktop + Go daemon — your repo never leaves your laptop unless you decide otherwise.
Roady
active · Go
Plan-of-record for AI coding agents. Spec, plan, and drift detection that survive context resets. File-based, git-versioned, MCP-native.
TokenOps
active · Go
Local MCP server + CLI that tracks Claude Max / ChatGPT Plus/Pro/Team / Copilot / Cursor plan windows and warns before you hit the cap. Output-side coach detects compressed reply styles.
Scout
active · Go
Browser automation in one binary. Simpler alternative to Playwright — no Node, no Python, no runtime. Library, CLI, MCP server, chat UI.
Fortify
active · Go
Production-grade resilience for Go services calling LLMs and tools. Circuit breaker, retry, rate limit, timeout, bulkhead, fallback, hedge, adaptive concurrency, cost budget, stream timeout. Zero core deps.
Bolt
active · Go
Zero-allocation slog.Handler for Go with first-class OpenTelemetry. Same encoder behind a chained-builder API for hot paths. Includes a genai add-on for the OTel GenAI semantic conventions.
Statekit
active · Go
Typed statechart engine for Go. Hierarchy, guards, actions, parallel states, snapshots, lint, visualization. Same primitives for backend domain workflows and AI agent runtimes.
Mnemos
active · Go
Evidence layer for AI agents — claims, contradictions, citations. Self-hostable. Wrappers in TypeScript and Python.
Nox
active · Go · Apache 2.0
Open-source security scanner with first-class AI app security: prompt injection at the call site, embedding leakage, agent over-privilege, MCP hardening, cross-file AI taint, polyglot AIBOM. Offline-first, agent-native via MCP. Lives at nox-hq.dev.
Cognitive stack
active · Go
Five composable Go libraries for AI agent runtimes: Mnemos (memory), Chronos (time/pattern perception), Nous (commitment extraction), Praxis (typed actions), Olymp (control plane). Domain-agnostic.
mcp-go
active · Go
Go framework for building MCP (Model Context Protocol) servers. Used by most of the projects above.
Govee Light Management
stable · TypeScript
Stream Deck plugin for managing Govee smart lights, with advanced group functionality. Local API, no cloud app required.
cclint
active · TypeScript
Fast, extensible linter for CLAUDE.md context files. Validates and optimises the rules an AI agent reads at session start.
Everything else (smaller utilities, abandoned experiments) lives at github.com/felixgeelhaar.
SPEAKING / TOPICS
Conference talks and recordings when the room is right — AI-agent infrastructure, agent-runtime primitives, deploy/release/launch decoupling, AI app security. The first recording lands here after the next conference; the long-form versions live in Field Notes. Topics I’m currently shaping into talks:
- An agent runtime is five primitives. Most fake at least three.
- Plan-of-record and evidence layer — the two memories your agent needs.
- Deploy. Release. Launch. Three calendars, three owners.
- The AI failure surface your security scanner can’t see.
Speaker invitations welcome — send the brief.
STUDIO / KLARLABS
Klarlabs
The Munich studio behind most of the shelf above. Product engineering for AI-agent infrastructure — its own site, sitting next to this one.
HIRE / CONTACT
Working on agent infrastructure, platform topology, or an AI feature that has to hold up in production? Email is the fastest path: hello@felixgeelhaar.com. One paragraph with the actual question beats a long brief.