LLM Client
@n-dx/llm-client is the vendor-neutral LLM foundation layer. It provides shared provider interfaces, Claude and Codex adapters, provider registry, and token usage tracking.
Role in the Architecture
LLM Client sits at the foundation tier — the lowest level of the dependency hierarchy. Both Rex and Hench depend on it for all LLM interactions.
hench → @n-dx/llm-client
rex → @n-dx/llm-clientWhat It Provides
- Provider interfaces — abstract types for LLM calls, responses, and token usage
- Claude adapter — API and CLI modes for Anthropic's Claude models
- Codex adapter — CLI adapter for OpenAI's Codex models
- Google adapter — API adapter for Google's Gemini models
- Local adapter — OpenAI-compatible adapter for a local LM Studio / Ollama server
- Provider registry — resolve the configured vendor to the right adapter
- Token usage tracking — unified token counting across providers
- Help formatting — shared terminal output formatting utilities
- JSON utilities — robust JSON extraction and repair for LLM responses
Vendor Support
| Vendor | API Mode | CLI Mode | Token Accounting | Default model |
|---|---|---|---|---|
| Claude | Yes (recommended) | Yes | Full | claude-sonnet-5 |
| Codex | No | Yes | Limited (CLI doesn't return usage) | gpt-5.6-terra |
| Yes | No | Full | gemini-2.5-pro | |
| Local | Yes (OpenAI-compatible) | No | Depends on server | (whatever is loaded) |
Defaults live in NEWEST_MODELS / TIER_MODELS in src/config.ts — that is the single place to update when a vendor ships a new model.
Not a Public API
LLM Client is an internal package (@n-dx/ scoped). Its API surface is consumed exclusively by Rex and Hench through their gateway modules. External consumers should use the n-dx CLI rather than importing this package directly.