Parity Overview
Claw Code Agent is a ground-up Python reimplementation of Claude Code’s npm-based agent architecture. The PARITY_CHECKLIST.md tracks implementation status across every feature surface. The project spans ~21,000 lines across 51 source files, achieving strong core parity while significant gaps remain in interactive features and infrastructure.
What’s Fully Implemented
Core Agent Runtime
| Feature | Status |
|---|---|
| One-shot agent loops | Complete |
| Iterative tool calling | Complete |
| Streaming output (token-by-token) | Complete |
| Local model execution (vLLM, Ollama, LiteLLM, OpenRouter) | Complete |
| Session persistence and resumption | Complete |
| Cost tracking and budget enforcement | Complete |
| File history journaling with snapshots | Complete |
| Nested agent delegation with topological batching | Complete |
| Context compaction (auto-snip and auto-compact) | Complete |
| Structured JSON schema output mode | Complete |
| Truncation continuation | Complete |
CLI and Commands
| Feature | Status |
|---|---|
Python CLI (python3 -m src.main) | Complete |
agent / agent-chat / agent-resume commands | Complete |
Background sessions (agent-bg, agent-ps, agent-logs, agent-attach, agent-kill) | Complete |
Diagnostic commands (summary, manifest, parity-audit, command-graph) | Complete |
| 37 slash commands (of ~80+ upstream) | Complete |
Prompt and Context
| Feature | Status |
|---|---|
| Structured system prompt builder (21 sections) | Complete |
| CLAUDE.md discovery (global, directory, rules) | Complete |
| Git state detection (branch, status, recent commits) | Complete |
| Tokenizer-aware context accounting | Complete |
| Environment snapshot (platform, shell, date) | Complete |
Tools
| Feature | Status |
|---|---|
| 58 built-in tools implemented | Complete |
| File operations (read, write, edit, list, glob, grep) | Complete |
| Shell execution with full security validation | Complete |
| MCP resource access and tool invocation | Complete |
| Account, config, search, remote management | Complete |
| Task/plan/team/workflow/worktree management | Complete |
| Nested agent delegation | Complete |
Security
| Feature | Status |
|---|---|
| Bash security (18 validators, 163 tests) | Complete |
| Permission tiers (read/write/shell/unsafe) | Complete |
| Hook policy manifests | Complete |
| Tool blocking (exact and prefix match) | Complete |
| Destructive command detection (20+ patterns) | Complete |
What’s Partially Implemented
CLI Modes (Partial)
Working: local background sessions, daemon wrapper, agent-chat interactive mode.
Missing: remote-control/bridge runtime mode (30+ files in npm), browser/native-host runtime, computer-use MCP mode, template jobs, environment runners.
Slash Commands (Partial)
Implemented (37 commands): /help, /context, /prompt, /permissions, /model, /tools, /memory, /status, /clear, /config, /account, /search, /mcp, /remote, /tasks, /plan, /hooks, and others.
Missing (43+ commands): /add-dir, /bridge, /chrome, /desktop, /ide, /install-github-app, /keybindings, /plugin, /release-notes, /sandbox-toggle, /theme, /upgrade, /voice, /vim, plus feature-gated and internal variants.
Tools (Partial)
Core tools have full fidelity. Gaps exist in:
| Missing Tool | Purpose |
|---|---|
AgentTool (full) | Sub-agent spawning with built-in agent types |
SkillTool | Skill loading and execution |
LSPTool | Language Server Protocol diagnostics |
PowerShellTool | PowerShell with security validators |
REPLTool | Interactive REPL execution |
MCPTool (full) | Complete MCP execution beyond stdio |
What’s Not Implemented
Remote and Team Infrastructure
The npm Claude Code implementation includes a bridge subsystem (30+ files) for real remote session management with WebSocket infrastructure, shared remote state, and upstream proxy. None of this is implemented — the Python version has local remote profile management only.
Interactive UI / REPL
The npm implementation includes an Ink/TUI framework with 40+ files, 100+ React-style components, a screen system, virtual scrolling, approval UI flows, keyboard interaction, and rich incremental rendering. The Python version provides functional CLI execution but no interactive TUI parity.
MCP / Plugin / Skill Services
Local MCP resource access works over stdio. Missing: full MCP service (25+ files covering auth, permissions, config, registry, OAuth), plugin discovery/loading/installation, bundled skill support, and plugin lifecycle management.
Services Layer
The npm implementation includes extensive backend services not present in Python:
| Service | Purpose |
|---|---|
| Analytics | Usage tracking and telemetry |
| LSP | Language Server Protocol integration |
| Auto-dream | Background task generation |
| Agent summary | Conversation summarization |
| Magic docs | Documentation generation |
| Session memory | Cross-session knowledge persistence |
| Prompt suggestion | Contextual prompt recommendations |
| Memory extraction | Automatic insight extraction |
| OAuth | Authentication flow management |
| Rate limiting | API throttling |
| Settings sync | Cross-device configuration |
Platform Integration
| Feature | Status |
|---|---|
| Voice mode | Not implemented |
| VIM mode | Not implemented |
| Keybinding system (13 files) | Not implemented |
| IDE integration (VS Code, JetBrains) | Not implemented |
| Notification hooks | Not implemented |
| Chrome extension | Not implemented |
Enterprise Features
| Feature | Status |
|---|---|
| Coordinator mode | Not implemented |
| Buddy/companion system (6 files) | Not implemented |
| Migration system (11 scripts) | Not implemented |
| Growthbook feature flags | Not implemented |
| Internal logging infrastructure | Not implemented |
Development Roadmap
The PARITY_CHECKLIST.md defines a tiered priority system:
Tier 1 — High Priority
- LSP tool integration for code intelligence
- Full
AgentToolwith built-in agent types (Explore, Plan, etc.) - Auto-compact and context collapse improvements
- Interactive REPL enhancements
Tier 2 — Medium Priority
SkillToolimplementation- Full MCP service beyond stdio transport
- Plugin discovery, loading, and installation
- Real remote session management
- Implementation of remaining 43+ slash commands
Tier 3 — Nice to Have
- TUI components and interactive rendering
- Voice and VIM modes
- IDE integrations (VS Code, JetBrains)
Tier 4 — Enterprise / Platform
- Bridge subsystem for remote sessions
- OAuth and rate limiting services
- Auto-dream and consolidation tasks
- Analytics and telemetry
Architecture Comparison
| Dimension | Claude Code (npm) | Claw Code Agent (Python) |
|---|---|---|
| Language | TypeScript/Node.js | Python 3.10+ |
| Dependencies | npm ecosystem | Zero (stdlib only) |
| Model backend | Anthropic API only | Any OpenAI-compatible endpoint |
| UI framework | Ink (React for CLI) | Plain stdout/stdin |
| Codebase size | ~100K+ lines | ~21K lines |
| Tools | 80+ | 60+ |
| Slash commands | 80+ | 37 |
| Source files | Hundreds | 51 |
| MCP transport | stdio, SSE, WebSocket | stdio only |
| Plugin system | Full lifecycle management | Manifest-based hooks and aliases |
| Remote sessions | WebSocket bridge (30+ files) | Local profile management |
| TUI | Ink components (100+) | None |
| Platform | macOS, Linux, Windows | macOS, Linux |
Key Design Differences
-
Zero dependencies — Where Claude Code relies on hundreds of npm packages, Claw Code Agent uses only Python’s standard library. This is a deliberate architectural choice for portability, auditability, and simplicity.
-
Backend agnosticism — The OpenAI-compatible client abstracts away model-specific details. Switching from vLLM to Ollama requires only changing three environment variables.
-
Immutable state patterns — The
AgentManageruses frozen dataclasses and reconstruction-based updates rather than mutation, providing stronger auditability guarantees. -
Reactive context management — Rather than proactively managing context, the agent reacts to prompt-length errors with snipping and compaction, then retries.
-
Functional CLI — No attempt to replicate the Ink-based TUI. The focus is on correctness of the agent loop rather than interactive aesthetics.