智能代理操作系统

Install dependencies

npm install

1. Install

npm install stoic-agentos-sdk

Quick Start

2. Get Your API Key

Sign up at stoic-agentos.vercel.app → Dashboard → Settings → Generate Key

API Reference

SDK Reference

import { AgentOS } from 'stoic-agentos-sdk';

// Initialize
const os = new AgentOS({ apiKey: 'sk_live_xxx', workspace: 'my-app' });

// Core methods
os.capture({ type, title, content, metadata })     // Log observation
os.wrapAgent(name, fn)                              // Auto-monitor function
os.addKnowledge({ name, summary, content })         // Persist knowledge
os.listAgents()                                     // Get all agents
os.listObservations({ limit, type })                // Query observations

// Claude-powered insights (v2.1+)
await os.summarize({ hours: 168 })                 // AI briefing of recent activity
await os.analyzeAgent(agentId)                     // Diagnose an agent's health
await os.ask('Why did the email-agent fail?')      // Free-form Q&A

Claude Integration

AgentOS uses Anthropic Claude for AI-powered insights — summarizing observations, diagnosing agent failures, answering free-form questions about your fleet.

Models: Haiku 4.5 for fast summaries, Sonnet 4.6 with adaptive thinking for deep diagnosis.

Three surfaces: - API: POST /insights/{summarize,analyze-agent,ask} — see API Reference - SDK: os.summarize(), os.analyzeAgent(id), os.ask(q) (above) - MCP server: agentos_summarize_observations, agentos_analyze_agent, agentos_ask tools

BYOK (Bring Your Own Key): Customers can route inference through their own Anthropic account from Settings → Anthropic API Key. Keys are stored encrypted in Supabase Vault (vault.secrets, pgsodium at rest) and accessed only by the API's service role. When no per-org key is set, the platform falls back to the ANTHROPIC_API_KEY env var.

Cost tracking: Every Claude call is logged to anthropic_usage with token counts and cache hits. The Settings tab shows call count, token usage, and estimated cost over a 7/30/90-day window.

Caching: All requests use cache_control: { type: 'ephemeral' } so repeated system prompts hit the prefix cache at ~10% of input cost.