The lab behind Familiar.
Open Horizon Labs builds Familiar's substrate: the harness, kernel, alignment skills, and operating philosophy that let a persistent AI operator work across real systems without losing judgment at the boundary.
Muness Castle
Twenty years in engineering and data leadership — Shopify, Zapier, JPMC, NerdWallet, Fullscript. He has built data organizations from scratch, led teams through hypergrowth, and spent most of that time watching smart people lose work to context collapse: the gap between what was decided and what actually got built.
He thinks about software the way a manufacturing engineer thinks about a production line — find the constraint, remove it, find the next one. Right now the constraint is context. AI can generate faster than anyone can think, but without memory, intent, and constraints baked in, speed turns into thrash. Familiar is the operating substrate built from that constraint.
He also shipped firmware for a custom Roon controller he built from scratch, with 260+ forum posts and 51 tagged releases, having never written firmware before. He does the work.
Drazen Urch
Cognitive psychology and statistics degree, turned data engineer, turned systems engineer. Worked at Zapier, Token Terminal, Nym Technologies before co-founding Open Horizon Labs. His open source Rust libraries have over 50 million combined downloads.
He was building practical agent tooling — supervisor LLMs, hooks that intercept model uncertainty and route it to a human's phone — before that was a recognizable category. The kernel running Familiar is his architecture. The security model, the extension system, the live-reload loop: all of it came from running it himself and hitting the edges.
He writes about intent, judgment, and what it actually takes for an AI system to be trustworthy rather than just capable. The distinction matters because a familiar with broad connection must still know when not to act.
Proof in running systems.
The lab's public work is not a separate offer. It is the trail of infrastructure and alignment practice that Familiar is built from.
Familiar
A persistent AI kernel that runs 24/7, writes its own tools, and shapes itself to the person or job running it. This site is managed by one.
superego
A metacognitive supervisor for AI coding agents. Hooks into Claude Code, evaluates approach before large edits, blocks and injects feedback on scope drift. GitHub →
Repo-native alignment
Business outcomes, signals, and constraints as queryable repo artifacts. Gives AI agents the context they need to stay aligned with what actually matters. GitHub →
open-horizons-oss
Self-hostable strategy graph for aligning work to outcomes. The Open Horizons framework as running software — intent, goals, and constraints queryable by humans and agents alike. GitHub →
HiPhi
A Roon-native control stack for high-end audio installs — custom ESP32 firmware, e-ink displays, AI-powered zone management. Built because the alternatives were broken. hiphi.audio →
oh-omp
The agent runtime powering Familiar. Unified model API across 50+ providers, LSP integration, browser control, subagents, and streaming. The execution layer everything else runs on. GitHub →
Bring us one operating loop that should not reset.
Familiar is for principals who need continuity across tools, people, and decisions while keeping authority narrow and visible.