@agentskit/core — for agents

#Purpose

Stable TypeScript contracts (Adapter, Tool, Memory, Retriever, Skill, Runtime) + createChatController + primitives. Target: <10 KB gzipped, zero runtime deps.

#Install

npm install @agentskit/core

#Primary exports (main entry)

• createChatController(config) — state machine behind every UI
• proposeToolCall(...) — place a trusted, validated action into the same confirmation lifecycle; never executes directly binding. Returns { getState, subscribe, send, stop, retry, edit, regenerate, setInput, clear, approve, deny, updateConfig }.
• defineTool(config) — typed tool factory with JSON-Schema input inference.
• ChatConfig.authorizeToolCall — trusted host authorization before tool proposal and again before execution. Callback failures default to deny.
• createInMemoryMemory / createLocalStorageMemory — default ChatMemory implementations.
• createStaticRetriever / formatRetrievedDocuments — RAG-less retrieval.
• executeToolCall, consumeStream, createEventEmitter, safeParseArgs, generateId, buildMessage — low-level helpers.
• AgentsKitError, AdapterError, ToolError, MemoryError, ConfigError, RuntimeError, SandboxError, SkillError, ErrorCodes — error taxonomy.
• Token budget: compileBudget, approximateCounter.
• Progressive tool args: createProgressiveArgParser, executeToolProgressively.
• createVirtualizedMemory — hot-window + cold-retriever wrapper (also re-exported from @agentskit/memory).
• Untrusted-input fencing lives in the @agentskit/core/security entry: fenceUntrustedContent(content, opts?) + UNTRUSTED_CONTENT_DIRECTIVE wrap attacker-influenced text so the model treats it as data, not instructions (the mitigation companion to createInjectionDetector).
• Multi-modal content parts: textPart, imagePart, audioPart, videoPart, filePart, partsToText, normalizeContent, filterParts.
• Agent-loop internals (advanced): buildToolMap, activateSkills, executeSafeTool, createToolLifecycle.
• Memory serialization: serializeMessages, deserializeMessages; validate untrusted records with validateMemoryRecord from @agentskit/core/memory-validation.

#Subpath exports (zero main-bundle weight)

#Minimal example

import { createChatController } from '@agentskit/core'
import { anthropic } from '@agentskit/adapters'

const controller = createChatController({
  adapter: anthropic({ apiKey: process.env.ANTHROPIC_API_KEY!, model: 'claude-sonnet-4-6' }),
})

controller.subscribe(() => console.log(controller.getState().messages))
await controller.send('Hello')

#Common patterns

• Wire a controller into any framework via getState + subscribe (see the framework bindings under For agents).
• Compose tools with composeTool or wrap failing ones with wrapToolWithSelfDebug.
• Gate risky operations with createApprovalGate + HITL + signed audit log.

#Related packages

• @agentskit/adapters
• @agentskit/runtime
• @agentskit/react

#Source

• npm: https://www.npmjs.com/package/@agentskit/core
• repo: https://github.com/AgentsKit-io/agentskit/tree/main/packages/core