@jambonz/llm

A voice-oriented LLM adapter library for Node.js.

@jambonz/llm provides a uniform, vendor-neutral interface for streaming chat completions from large language models — with first-class support for tool calls, AbortSignal-based interruption, and lossless conversation-history round-tripping. It is the LLM layer that powers jambonz voice agents, but the library itself has no voice-specific dependencies and can be used from any Node.js application.

Design goals

1. A small, well-defined interface for adding new vendors. Implement one adapter class plus a manifest, register it, and the contract test-kit tells you whether you're done.
2. Caller-owned conversation history. The library is stateless across calls. Barge-in, history trimming, speculative preflight, and other voice-specific concerns are implemented by the caller, not the library.
3. Lossless history round-trip. Vendor-native message shapes (Anthropic content blocks, OpenAI tool_calls arrays, Google parts, Bedrock content[]) are preserved via an opaque vendorRaw escape hatch.
4. Community-friendly contribution. A new vendor is a single directory under src/adapters/, plus a line in src/adapters/index.ts.

Contributing

New vendor adapters are the primary contribution path. See docs/adding-a-vendor.md for the step-by-step guide — copy the template, fill in the adapter, run the contract test-kit. A PR is merge-ready when all contract-kit checks pass against the new adapter and the build is green.

Status

Under active development. API may change before 1.0.0.

License

MIT