llm-corePortable contracts for AI delivery and agent runtimes

Govern intent, authority, evidence, and runtime interoperability without hiding a home-grown execution engine.

Get started

Understand the boundary

Runtimes stay native

AgentRunner is implemented by qualified integrations; native graphs, sessions, and checkpoints remain runtime-owned.

Delivery stays governed

Specifications, authority, evaluation, provenance, and evidence connect AI-first delivery to the software it produces.

Effects follow one path

Policy, approval, execution, and receipts remain distinct, so consequential effects fail closed.

Interoperability is qualified

Versioned adapters declare capabilities and semantic loss instead of promising universal behavior.

Install

The llm-core 2 candidate is ESM-only and requires Node.js 22 or newer. It is not yet published to npm.

bash

# Available after the 2.0.0 release has registry evidence
npm install @geekist/[email protected]

The short version

llm-core gives AIFSD and other hosts one portable contract for describing AI intent, qualifying implementations, controlling effects, and retaining evidence. It is not an agent SDK or workflow engine. It serves:

AI-first software delivery tools that understand, specify, build, review, evaluate, approve, and release software; and
qualified runtime integrations that execute agentic behavior inside the delivered product.

The package does not select a default runner. A host or runtime integration implements AgentRunner and supplies it explicitly. Concrete runtime adapter subpaths become supported only after their own qualification and publication.

Choose the route that matches your work:

Portable capabilities for contract users.
Agents and runtime integrations and qualified adapters for implementers.
Packaging and conformance and evidence for qualification consumers.
@aifsd/sdk for application composition. Its current candidate supports only config and integrations; application composition remains planned.

Package exports lists every public front and its owner.