Claude Code Leak Exposes Models & Agent Features
Anthropic's 500k-line Claude Code leak reveals codenames for Opus (Fenick), Sonnet (Capra), upcoming Opus 4.7/Sonnet 4.8, Mythos with 1M context, and 44 feature flags like multi-agent coordination and infinite memory.
Developer Recreations Unlock Claude Code Locally
The leak of over 500k lines of Claude Code source code—accidentally shipped via a GitHub debugging file—exposed Anthropic's full agent architecture in TypeScript, including strict permission controls and token-optimized designs. Developers quickly mirrored it, making takedowns futile. One recreated it as a self-improving system: compiled locally, analyzed by an agent team using open models, stripping telemetry, guardrails, and unlocking experiments like ultra plan for long-running tasks and async multi-agent research. Another built 'free code' CLI with background sessions, removing restrictions for independent operation. This blueprint lets builders clone and modify Claude's agentic coding system today, bypassing API limits for local experimentation.
Model Codenames and Internal Benchmarks
Files confirm codenames: Fenick for Opus series (tested earlier this year, excelled at building full browser-based OS across Windows/Mac with frontend/backend/agentic components); Tangu for Haiku; Capra for Sonnet (hints at 1M context version); Titan. Internal tests reference Opus 4.7 and Sonnet 4.8 already powering Claude Code, plus Mythos (next flagship tier) and Capra with 1M token context, fast/full reasoning modes. Kappa Barrett mentioned vaguely. These run behind hidden flags, signaling releases soon—Opus lineage showed superior multi-domain coding, like complete OS UIs from scratch.
44 Feature Flags Reveal Agent Advancements
Exposed capabilities fix Claude Code's limits: autodream compresses session history for infinite memory; Chyros (proactive agent wakes itself for tasks); Karios (background AI); coordinator spawns parallel workers; voice mode for real-time talks; advisor (server-side stronger model oversees conversations via API for quality); undercover mode hides AI identity; structured long-term memory; dispatch for remote control; MCP for full computer control; peer discovery between instances; bug hunting, session teleportation, AFK autonomous ops. Buddy is an April Fool's Tamagotchi-style pet for fun feedback. Multi-agent coordination and ultra plan enable extended research sessions, turning Claude into a daemon-like long-runner—directly addressing memory and independence issues for production agent workflows.