AgentSky Guide 2026: Run Claude Code, Codex & Hermes in the Cloud with One API (The "OpenRouter for Agents")

"Claude Code, Codex, Hermes — what if you could run them all in the cloud with a single API key?"

The short version: AgentSky is an "OpenRouter for agents" — a cloud API platform that runs 8 AI agents (Claude Code, Codex, Hermes, pi, opencode, Kimi Code, OpenClaw, DeepSeek Harness) with one API key. You pick the agent and the model freely, connect channels like Slack, Telegram, Discord, iMessage, or your own API, and persistent sessions keep state between messages. It starts with $3 free credit, no card required, and if you connect your Claude Pro or ChatGPT subscription, eligible agents run at $0 model usage. As of August 2026 it was Product Hunt's Product of the Day with 10K+ sessions in production.

This article explains what AgentSky is, what you can do with it, pricing, how to use it, comparisons, and caveats — based on official information.

What AgentSky Is

AgentSky (https://agentsky.dev) is a cloud API platform for AI agents. The official site calls itself "The OpenRouter for agents."

While OpenRouter gives you access to many LLMs through one API, AgentSky lets you run many AI agents through one API. OpenAI's GPT API calls a model; AgentSky calls an agent to get work done.

Supported agents (8):

Claude Code
Codex
Hermes
pi
DeepSeek Harness
opencode
Kimi Code
OpenClaw

Supported models (Model Router):

GPT-5.6 Sol / GPT-5.6 Terra
Claude Opus 5 / Claude Fable 5 / Claude Sonnet 4.6
Gemini 3.5 Flash
DeepSeek V4 Pro / DeepSeek V4 Flash
Kimi K3
GLM-5.2
Grok 4.6

Agent and model are freely combinable. The key idea: decouple the harness (agent) from the model — e.g. "Claude Code + DeepSeek V4 Flash" or "Codex + GPT-5.6 Sol."

What You Can Do

1. Move your local agent to the cloud in one command

sky clone

This clones your local agent — same instructions, same model, same MCP servers — into a cloud agent that keeps working while your laptop is closed.

2. Connect channels and delegate via chat

Connect Slack, Telegram, Discord, WhatsApp, iMessage, Web, API, or A2A protocol, and delegate work through chat:

Ask in Slack: "Fix today's bug" → Codex works on the repo
Ask in Telegram: "Write a competitive analysis report" → Hermes researches and writes
Embed in your own app's API to offer AI agents to your users

3. Use tools & connectors on the same API call

Built-in tools include search & web fetch, browser, SEO & SERP data, image generation, background removal, video generation, transcription. Add app connectors like Gmail, Notion, Slack — 2,000+ apps. All specified as parameters on the agent call — nothing separate to set up.

4. Embed agents in your product (B2B)

Customer stories include:

Tycoon: an "AI CEO" routes tasks to Claude Code, Codex, and Hermes through one API (150,000 agents launched)
HeyBoss AI: pre-warms cloud agents to eliminate cold-start delay (250,000 agents launched)
WebJourney: moved from Claude Code to pi + open-source models with one config change (50,000+ agents launched)

Agent Arena: Which Combo Is Best?

AgentSky publishes a benchmark "Agent Arena" based on real production data (40,835 tasks, $141,676 measured spend, May–Aug 2026). These are Elo ratings from head-to-head wins on real user tasks — not synthetic suites.

#	Harness × model	Elo	Win rate	Cost/task	Time/task
1	Claude Code × Claude Fable 5	1,201	60.6%	$4.86	6m57s
2	Codex × GPT-5.6 Sol	1,185	58.4%	$1.88	5m54s
3	Claude Code × Claude Opus 5	1,170	56.2%	$3.63	7m40s
4	Claude Code × Claude Opus 4.8	1,130	50.5%	$2.20	5m54s
5	Hermes × DeepSeek V4 Pro	1,104	46.8%	$0.53	6m10s
6	Claude Code × Claude Sonnet 4.6	1,090	44.8%	$1.52	5m26s
7	Hermes × DeepSeek V4 Flash	1,068	41.7%	$0.14	2m29s
8	Hermes × Kimi K3	1,063	41.0%	$2.26	10m26s

How to read it:

Quality first: Claude Code × Claude Fable 5 (Elo 1,201, 60.6% win)
Value first: Hermes × DeepSeek V4 Flash ($0.14/task, 2m29s) or Hermes × DeepSeek V4 Pro ($0.53/task)
Balanced: Codex × GPT-5.6 Sol ($1.88/task, Elo 1,185)

Elo centers on 1000; a 40-point gap is a clear edge, 10 points is noise. Boards are broken down by category (coding, research, support, marketing, sales, email, data, SEO, content).

Pricing

Two mechanisms:

1. Platform billing (pay-as-you-go)

Per-minute agent runtime + per-token model usage
$3 free credit, no card, no setup
Parked (idle) agents are free

2. Bring Your Own Subscription

Connect Claude Pro / Claude Max / ChatGPT and eligible agents run at $0 model usage:

Claude: run claude setup-token (Anthropic's official headless sign-in), paste into Settings
ChatGPT: sky auth connect chatgpt reads your Codex login
Shows "Runs on your Claude subscription — $0 model usage"
Only runtime (machine time) remains on the platform side
If your plan doesn't cover a model, the turn fails — your wallet is never silently charged

How to Use (3 Steps)

Sign up ($3 free credit, no card)
Get your API key (AGENTSKY_API_KEY)
Name the agent and model in the request

{
  "agent": "claude-code",
  "model": "claude-fable-5"
}

Persistent sessions keep state (memory, files) between messages.

Comparison

Service	Focus	vs AgentSky
OpenRouter	Call LLMs via one API	AgentSky calls agents (model + harness)
Vercel AI SDK	Frontend AI integration	Dev SDK; AgentSky is an agent execution platform
LangChain	LLM app framework	Build your own agents; AgentSky runs ready agents
Vendor APIs (Anthropic/OpenAI)	Model APIs only	AgentSky bundles agent + model + channels + tools

Caveats

Can cost more than local: always-on cloud agents bill machine time; for quick checks, local may be cheaper
Subscription constraints: models and rate limits follow your plan; out-of-scope turns fail (no silent charge)
Quality depends on combo: Arena data is Elo from real tasks; combos under 250 tasks are provisional
Persistent cloud execution: review data handling/security policies before processing sensitive info
Information as of August 2026; pricing and models can change

Summary

AgentSky = "OpenRouter for agents": 8 agents × 12+ models via one API
Connect channels (Slack, Telegram, Discord, API...) to delegate work
2,000+ app connectors and built-in tools on the same API
$3 free credit, no card; connect Claude Pro/ChatGPT for $0 model usage
Agent Arena gives real quality-vs-cost data to pick your combo
sky clone moves your local agent to the cloud — works while your laptop is closed

The question of "where to run agents" has more options in 2026. AgentSky is one of the best fits if you want "one API for everything."

Sources

AgentSky official: https://agentsky.dev/
Agent Arena: https://agentsky.dev/benchmarks
Bring Your Own Subscription: https://agentsky.dev/tutorials/run-agents-on-your-claude-or-chatgpt-subscription
Product Hunt: https://www.producthunt.com/products/agentsky