CatCode — open-source coding harness that actually compounds.
Most agents forget everything between sessions and dump one giant turn at a problem. CatCode is different:
🧠 Self-learning built in — reflects mid-session, saves durable memories/skills, and injects them next time. /index a new repo once; it keeps getting smarter on your codebase.
🐶 Watchdog advisor — optional second model reviews the executor's work (correctness/security/tests), posts structured findings, and can soft-continue the turn. Fail-open: never blocks the agent if review fails.
🎯 Goal mode + Control Center — not "hope the agent finishes." Plan → (self-review) → deploy parallel subagents → verify evidence → replan until certified. CEO mode runs the loop without mid-mission babysitting.
🤝 Real multi-agent system — scout / planner / worker / reviewer with fork, parallel, chain, peer intercom, and optional git worktree isolation.
🔒 MicroVM sandbox — not a denylist. Real Microsandbox isolation (Linux KVM / Apple Silicon / Windows WHP), fail-closed.
🖥️ Hub that doesn't die when you close the tab — multi-project, multi-session chat; leave, sign out, or switch devices and reattach live cores mid-turn. Git + preview + terminal around the chat.
🔌 Extend without recompiling — hooks, custom tools, OAuth providers, prompt-backed slash commands (e.g. /deep-research).
Same Rust core under both terminal and browser. Open source.
🔗 https://github.com/catalystctl/catcode 🌐 https://code.catalystctl.com
Short version
CatCode — open-source coding harness that actually compounds.
• Learns your repo across sessions (memory + skills +
/index/reflect) • Watchdog second-model review on finished turns • Goal / Control Center: plan → parallel subagents → verify → replan until done • Real microVM sandbox, not pretend isolation • Multi-project Hub: close the browser, work keeps running — reattach anytimeTerminal + browser. One Rust core. Your models, your machine. https://github.com/catalystctl/catcode