Community-driven innovation on top of the Claude Code architecture.
This project takes the publicly exposed Claude Code source snapshot (leaked via npm source map on 2026-03-31) and uses it as a foundation for open research, experimentation, and innovation in agentic CLI tooling.
We are not claiming ownership of Anthropic's code. This is an educational and research fork focused on pushing the boundaries of what developer-facing AI agents can do.
| # | Innovation | Status | Description |
|---|---|---|---|
| 1 | Progressive Trust Escalation | Planned | Behavioral permission scoring that learns per-workspace trust over time |
| 2 | Predictive Context Management | Planned | Proactive compaction with semantic priority ranking instead of reactive truncation |
| 3 | Agent Mesh with Shared Knowledge | Planned | Multi-agent shared knowledge graph, conflict resolution, dynamic rebalancing |
| 4 | Hybrid Local/Cloud Model Routing | Planned | Route simple tasks to local models, complex reasoning to cloud APIs |
| 5 | Tool Failure Feedback Loop | Planned | Adaptive prompting based on recent tool execution telemetry |
| 6 | Structured Episodic Memory | Planned | Vector-indexed memory with cross-project transfer learning |
| 7 | Real-time Collaboration | Planned | Multi-user shared Claude sessions |
See ARCHITECTURE_REPORT.md for the full technical deep-dive of the original codebase.
- ~1,900 files, 512,000+ lines of TypeScript
- Runtime: Bun
- Terminal UI: React + Ink (Yoga flexbox)
- CLI: Commander.js
- Validation: Zod v4
- Protocols: MCP, LSP
- API: Anthropic SDK
- Feature Flags: GrowthBook (compile-time via
bun:bundle+ runtime)
src/
query.ts # Main query loop (async generator, tool execution, compaction)
QueryEngine.ts # LLM orchestration engine
Tool.ts # Tool type system
tools/ # 40+ tool implementations
commands/ # 50+ slash commands
components/ # 140+ Ink/React UI components
bridge/ # IDE integration (VS Code, JetBrains)
coordinator/ # Multi-agent orchestration
services/mcp/ # Model Context Protocol
plugins/ # Plugin architecture
skills/ # Reusable workflow system
memdir/ # Persistent memory
hooks/ # Permission system + React hooks
vim/ # Full vim mode state machine
voice/ # Voice input (STT)
buddy/ # Companion sprite system
remote/ # Remote CCR sessions
server/ # Server mode
tasks/ # 7 task types
keybindings/ # Chord-based keybinding system
This is an open research project. We welcome:
- Architecture analysis and documentation
- Proof-of-concept implementations of roadmap items
- Security research and vulnerability analysis
- Performance profiling and optimization ideas
- Novel agent interaction patterns
- The original Claude Code source is the property of Anthropic.
- This repository is not affiliated with, endorsed by, or maintained by Anthropic.
- This project exists for educational and research purposes.
- The source was publicly accessible via a source map exposure in the npm distribution.
The original source code is proprietary to Anthropic. Innovations and additions by this project's contributors are shared for educational and research purposes under fair use.