DartClaw

Security-hardened AI agent runtime for Dart. A single AOT-compiled Dart binary orchestrates Claude via a 2-layer architecture: Dart host (state, API, security) spawns the native claude CLI binary as a subprocess, communicating via bidirectional JSONL over stdin/stdout.

Status: Pre-alpha — API is unstable and will change. This is an early development release to establish pub.dev presence. Real release planned for 0.4.

Architecture

┌─────────────────────────────┐
│  Dart Host (AOT-compiled)   │
│  ┌───────────┐ ┌──────────┐ │
│  │ GuardChain│ │ Sessions │ │
│  └─────┬─────┘ └──────────┘ │
│        │ JSONL stdin/stdout │
│  ┌─────▼─────────────────┐  │
│  │  claude CLI (Bun bin) │  │
│  └───────────────────────┘  │
└─────────────────────────────┘

Zero npm/Node.js at runtime. The claude binary is a self-contained Bun standalone executable. Dart handles orchestration, security policy, session persistence, and multi-channel messaging.

Core abstractions

  • AgentHarness — subprocess lifecycle, turn execution, event streaming
  • Guard / GuardChain — security policy evaluation (command, file, network, content guards)
  • Channel — messaging interface (WhatsApp, Signal)
  • BridgeEvent — sealed event hierarchy from the JSONL control protocol
  • DartclawConfig — YAML-based configuration

What this enables

  • Custom agent CLI/server apps using DartClaw's harness and guard chain
  • Embedding agent capabilities into existing Dart applications
  • Building alternative UIs (Flutter, CLI REPL) on top of core services
  • Reusing guard infrastructure for other agent runtimes

Current status

This package re-exports the core interfaces from the internal dartclaw_core package. The API surface is intentionally narrow for this pre-alpha release — only stable abstractions and models are exposed.

See the repository for development progress, roadmap, and documentation.

License

MIT

Libraries

dartclaw
DartClaw — Security-hardened AI agent runtime for Dart.