NLA

Natural Language Adapters.

NLA is a protocol and SDK direction for wrapping software surfaces behind a consistent natural-language-friendly interface.

The core idea is simple:

Take any surface that accepts structured input and can produce structured output.
Wrap it in a small adapter.
Expose it through a standard protocol.

That surface might be:

• a CLI
• an HTTP API
• an MCP server
• an SDK or library
• a local app automation surface
• a long-lived agent runtime

Thesis

Most "agent integrations" are really adapter problems.

The hard part is usually not the model. The hard part is turning an existing system into something that can:

• declare what it does
• accept validated input
• stream structured progress
• ask for missing input or approval
• return typed output and artifacts
• behave consistently across hosts

NLA aims to standardize that boundary.

Design Goals

• Make simple adapters very easy to write.
• Keep the protocol transport-agnostic.
• Support one-shot invocation and long-lived sessions.
• Preserve structured I/O instead of collapsing everything into plain text.
• Make side effects, approvals, and risk explicit.
• Work for both native NLA providers and foreign adapters around existing tools.

Non-Goals

• NLA is not an LLM provider.
• NLA is not a workflow engine.
• NLA is not a UI framework.
• NLA is not limited to coding agents.
• NLA does not require a specific host application.

Initial Docs

• PROTOCOL.md: first-pass nla/v1 protocol draft

Workspace

Current packages:

• @nla/protocol: core TypeScript types for the draft nla/v1 envelope and profiles, plus message-shape validation helpers
• @nla/host-core: generic host-side session client primitives for talking to long-lived NLA providers
• @nla/sdk-core: a small adapter runtime with handler-based invoke and session flows, including operation input/output schema validation
• @nla/sdk-http: an outbound HTTP adapter layer for wrapping APIs behind NLA operations
• @nla/sdk-cli: a thin CLI wrapper layer on top of sdk-core
• @nla/test: a lightweight adapter/runtime harness for tests and local verification
• @nla/transport-http: an HTTP transport for host-adapter communication using streamed NDJSON responses
• @nla/transport-stdio-jsonl: a stdio JSONL transport with live streaming flush support and a Node child process JSONL transport helper for hosts

Planned Package Shape

The repo starts as protocol-first documentation. If the shape holds up, the next step is likely a layered package structure:

packages/
  protocol/
  host-core/
  sdk-core/
  sdk-cli/
  sdk-http/
  transport-http/
  sdk-mcp/
  test/
examples/

Working Vocabulary

• NLA Protocol: the transport-independent contract
• NLA Provider: something that speaks NLA directly
• Foreign Adapter: a wrapper that translates another surface into NLA
• Host: the application or runtime that launches and talks to adapters

Status

Draft. This repo is the starting point for formalizing the protocol and then extracting a standalone SDK around it.

Local Development

npm install
npm run build
npm run typecheck
npm test
npm run example:http-transport-server
npm run example:stdio-echo
npm run example:stdio-cli
npm run example:stdio-http
npm run example:stdio-openrouter
npm run example:stdio-openrouter-adapter