agent-core

What it does

An agent that talks to OpenAI, Ollama, llama.cpp, vLLM or OpenRouter through the same API ends up writing the same plumbing: pooled clients, retries, streaming, and working out what each server refuses. This package is that plumbing, extracted from three servers that had each written it separately and drifted apart. Prompts, tools and what a run is for stay with the caller.

• Tool schemas a server accepts

  sanitizeTools makes an MCP tool schema something a strict or grammar-constrained server will take, and negotiate re-sends a request without what the endpoint or the model refused.

• Tools loaded on demand

  A name-only catalogue plus a load_tools meta-tool, so a run pays for the schemas the model asks for instead of all of them on every request.

• One turn, or the whole loop

  runTurn is one streamed turn with retry and negotiation around it; runAgentLoop runs it per step, with the tool calls between, until the model stops asking.

Get started

npm install @cubicecho/agent-core openai

import { runAgentLoop } from "@cubicecho/agent-core";

const { turn } = await runAgentLoop({
  config: {
    baseUrl: "http://localhost:11434/v1", // any OpenAI-compatible endpoint
    apiKey: "",
    model: "qwen3",
    maxTokens: 1024,
    temperature: 0.2,
    maxToolIterations: 8,
  },
  messages: [{ role: "user", content: "What changed in the last release?" }],
  tools, // OpenAI tool definitions, if the run has any
  dispatch: async ({ name, args }) => callTool(name, args), // returns what the model reads
});

console.log(turn.content);

Full documentation lives in the project README: https://github.com/cubicecho/agent-core