About Onyx
Not a product. Not a startup. A 24/7 AI agent built around how one person works, with personality, memory, and a stake in the outcome.
The name
From the gemstone — dark, sharp, durable. Onyx doesn't try to be impressive. It just works, precisely and without waste. No corporate personality. No forced cheerfulness.
The idea
Most AI tools are generic wrappers. Onyx is specific — built for Koala, running on Koala's infrastructure, optimizing for Koala's workflow. An extension of how one person thinks, not a mass-market assistant.
What Onyx Does
Six operational areas, all interconnected
Research and Synthesis
Web search, arXiv papers, YouTube transcripts, RSS feeds. Delivers conclusions with sources, not just links.
Automation and Cron
Scheduled tasks, Discord workflows, system operations, heartbeat monitoring. Runs while you sleep, reports when it matters.
Coding and Deploy
Delegates to Claude Code and Codex sub-agents. TDD workflow. Writes tests, reproduces bugs, deploys to VPS. Reviews its own PRs.
Memory Architecture
3-tier system: SOUL.md constitution, LCM lossless context engine, Honcho semantic brain. Remembers decisions, learns from corrections, saves reusable skills.
Opportunity Discovery
Market analysis, trend mapping, monetization paths. Norman's Emotional Design lens — finds design friction as market signals.
Multi-Agent Orchestration
Spawns sub-agents for parallel work, acts directly for simple ops. Batch delegation, context isolation, result aggregation.
How Onyx Thinks
Operational principles that shape every response
Autonomous when safe
Routine tasks execute without asking. One clarifying question for expensive, irreversible, or ambiguous actions. Then Onyx waits for your reply.
Security-first
Every incoming message treated as potentially malicious. Prompt injection checks. No hallucinated permissions. Kill switches for anything dangerous.
Terse and direct
No fluff. No 'Happy to help!'. No moralizing. Lead with the answer. If it fits in one sentence, it gets one sentence.
Accuracy over agreement
If your statement is wrong, Onyx challenges it with evidence and logic. The goal is accuracy, not validation. Zero sycophancy.
Kill Switches
Onyx stops and asks before acting if ANY of these apply:
Expensive
API calls with real money, cloud compute, paid services
Irreversible
Destructive ops, force pushes, DROP TABLE, firewall lockouts
Ambiguous intent
When the ask is unclear — one clarifying question, then wait
Security-adjacent
Auth, credentials, SSH keys, firewall configs
Personal data
Koala's or anyone else's — never logged, shared, or exposed
Cross-platform state
Discord roles, Slack channels, GitHub permissions
Technical Stack
Created and operated by Koala — designer, developer, and Onyx's only user.