@nazehs/agent-runtime
v0.1.2
Published
Framework-agnostic agent execution engine for Node (NestJS, Next.js API routes, workers).
Maintainers
Readme
@nazehs/agent-runtime
Local agent execution: register handlers, call execute / executeStream, get AgentResult. Optional guardrails, tracing, and timeouts are built in.
Installation
npm install @nazehs/agent-runtimeyarn add @nazehs/agent-runtimeUsage
import { createAgentRuntime, InMemoryAgentRegistry } from '@nazehs/agent-runtime';
import type { AgentContext, AgentResult } from '@nazehs/agent-runtime';
const registry = new InMemoryAgentRegistry();
registry.register({
name: 'my-agent',
version: '1.0.0',
executionMode: 'local',
handler: myAgent,
timeoutMs: 45_000,
});
const runtime = createAgentRuntime({
registry,
transportAdapters: [],
defaultTimeoutMs: 45_000,
});
const context: AgentContext = { requestId: crypto.randomUUID() };
const result: AgentResult = await runtime.execute('my-agent', input, context);Public types and functions are re-exported from the package entry; see the TypeScript definitions in dist/ or source index.ts.
Requirements
- Node
>= 20.10 - ESM only (
"type": "module"in your app, or dynamicimport()from CommonJS)
RAG
RAG (retrieval-augmented generation) is not built into the runtime; you do it in your handler—retrieve documents or snippets, add them to the system or user message list (or a dedicated RAG field your LLM client understands), then call the model. The example support path does this by grounding the system message from a local knowledge base (example/lib/support/support-llm-prep.ts).
Tool calling
Types — AgentToolDefinition, AgentToolCall, AgentToolResult, AgentInputWithMessagesAndTools. Context — optional context.tools to list tools for the run. Traces — emitAgentToolRequested / emitAgentToolCompleted / emitAgentToolFailed, and withAgentToolCallTraced for a single call. The LLM and tool↔message loop are implemented in your code; see example/lib/support for a Chat Completions tools / tool_calls loop and lookup_rentamnow_knowledge.
More
Full app example, env, git/tarball installs, and release process for this package live in the monorepo: github.com/Nazehs/agent-engine (see the root README and example/).
License
MIT — see LICENSE.
