What is this?

@kilnai/core is the contract and domain package behind Kiln, a control plane for governed AI work. It provides:

• Agent, capability, workflow, memory, task, channel, trigger, routing, and app contracts
• Provider adapter boundaries for direct and executable provider routes
• Governed execution, approval, authority, and verification contracts
• MCP client (Streamable HTTP) for external tool integration
• Memory Lattice governed memory records with scopes, layers, provenance, revisions, relations, lifecycle policy, recall, and bounded resource projection
• Knowledge (RAG) with chunkers, embedding adapters, and retrieval pipeline
• Safety pipeline: PII detection (6 types), content classification (6 categories), 4 policy rails
• Eval framework: 12 scorer types, YAML-configured experiments
• Typed error codes with context-aware suggestions

Install

bun add @kilnai/core

Usage

Define your app in YAML:

name: my-agent
runtime: provider-adapter
channels: [web]

provider:
  name: anthropic
  model: claude-haiku-4-5-20251001
  apiKeyEnv: ANTHROPIC_API_KEY

memory:
  scopes: [user]
  backend: sqlite+fts5

router:
  fallback: main

teams:
  main:
    agents:
      assistant:
        name: Assistant
        role: Helpful AI assistant
        goal: Answer questions clearly
        tier: fast
        tools: []
    workflow:
      phases: [respond]
    capabilities: []
    qualityGates: []

Load and validate it programmatically:

import { Engine } from "@kilnai/core";

const app = Engine.parseAppYaml(yamlString);
Engine.validateApp(app);

Key exports

| Namespace | Purpose | |-----------|---------| | Engine | 7 primitives, 3 composites, YAML loader, error catalog | | Orchestrator | Phase machine, checkpoint/resume, strategies | | Agents | Provider adapters, MCP client, tool cache, circuit breaker | | Memory | Governed Memory Lattice records, repository-backed persistence, SQLite adapter, lifecycle policy, recall, resources | | Safety | PII scanner, content classifier, policy rails, pipeline | | Knowledge | Chunkers, embedding adapters, vector store, retrieval | | Eval | 12 scorers, dataset loader, experiment runner | | Events | EventBus (32 typed events), EventStore | | Security | Audit log, prompt injection detection, secrets | | Cost | Per-role, cache-aware cost tracking |

Documentation

• Getting Started
• Core Concepts
• App Configuration
• Architecture

License

Apache-2.0