# AgentSky

> Run Claude Code, Codex, Hermes, OpenClaw, Pi, and more in the cloud through one API. Start an agent now or build agents into your product.

AgentSky (agentsky.dev) runs coding agents and their cloud computers through one API. Pick a supported harness such as Claude Code, Codex, Hermes, or OpenClaw, choose a compatible model, and use the agent directly or behind your product with managed state, tools, connectors, channels, and recovery.

## Product

- [Home](https://agentsky.dev/): what AgentSky is and how one-click launch works
- [Pricing](https://agentsky.dev/pricing): how billing works — prepaid credits, per-turn model usage, per-second compute while awake, free parking
- [Compare approaches](https://agentsky.dev/compare): model APIs, agent SDKs, DIY runtimes, and a complete cloud agent
- [For developers](https://agentsky.dev/developers): use coding agents as a hosted product backend
- [For software factories](https://agentsky.dev/software-factories): run coding workflows through one agent API
- [Agent Playground](https://agentsky.dev/benchmarks): harness + model rankings across quality, time, and cost
- [FAQ](https://agentsky.dev/#faq): common questions about harnesses, channels, pricing, data retention, and lock-in

## Use cases

- [Agent evaluation](https://agentsky.dev/use-cases/eval): Find the right agent and model for the job.
- [Long-running development](https://agentsky.dev/use-cases/long-running-development): Give the long job to a cloud coding agent.
- [Code review workflows](https://agentsky.dev/use-cases/code-review): Put a coding agent behind every review workflow.
- [Agent skill testing](https://agentsky.dev/use-cases/skills): Test one skill across different agent stacks.
- [Background agents](https://agentsky.dev/use-cases/background-agents): Keep an agent working after you close the laptop.
- [Agents in Slack](https://agentsky.dev/use-cases/slack-agent): Give your team an agent where it already talks.
- [Parallel task fan-out](https://agentsky.dev/use-cases/parallel-tasks): Fan one backlog out to many cloud agents.
- [Agent Builder](https://agentsky.dev/use-cases/launch-your-own-agent): Launch your own agent.

## Case studies

- [Tycoon developer architecture](https://agentsky.dev/case-studies/tycoon): Tycoon owns the product while AgentSky runs its coding agents and cloud computers

## Tutorials

- [Run cloud agents on the Claude or ChatGPT plan you already pay for](https://agentsky.dev/tutorials/run-agents-on-your-claude-or-chatgpt-subscription): Connect your Claude Pro/Max subscription or ChatGPT plan once and eligible agents run on it at $0 model usage — no per-token metering, no second bill. Setup takes two minutes in Settings or one CLI command.
- [How to benchmark your skills in 2026](https://agentsky.dev/tutorials/how-to-benchmark-your-skills-in-2026): A practical playbook for benchmarking agent skills — harness × model combinations — on your real work instead of synthetic leaderboards. Build the matrix, hold conditions identical, and pick a winner in an afternoon.

## FAQ

### What is AgentSky?

AgentSky is the OpenRouter for agents: one API that runs Claude Code, Codex, Hermes, pi, and more in the cloud. You pick the agent and the model it runs; the sandbox, tools, connectors, and channels ship as add-ons on the same call.

### Which agents and models can I run?

Eight agents run today: Claude Code, Codex, Hermes, OpenClaw, pi, DeepSeek Harness, Kimi Code, and opencode. Each pairs with compatible models across Claude, GPT, Gemini, DeepSeek, Kimi, and GLM — you pick both in the same call, and the full compatibility matrix is in the docs.

### How do I switch agents or models?

Change one word. “agent”: “claude-code” becomes “codex”; the model is a parameter the same way. Same request, same key, same bill. An existing agent keeps its engine — switching means a new run on a new agent, with your integration unchanged.

### I already run Claude Code locally — can I bring it to the cloud?

Yes. Clone to cloud recreates the Claude Code, Codex, or OpenClaw you already run as an always-on cloud agent: paste one prompt into it (or run sky clone) and it copies your instructions, model, and MCP servers. Config only — secrets, API keys, and session history never leave your machine. Eligible agents can even run on your existing Claude or ChatGPT subscription.

### What are the add-ons?

Everything beyond agent + model: a cloud computer with a persistent workspace, built-in tools like web search, browser, and image or video generation, 2,000+ app connectors with managed OAuth, and every channel — Web, API, Slack, Telegram, Discord, WhatsApp, iMessage, CLI, A2A. They are parameters on the same call, not separate products.

### How much does it cost?

New accounts receive $3 in free credit with no card required. You pay for usage; parked agents do not incur compute charges, and eligible agents can use a connected Claude or ChatGPT subscription at $0 model usage. The current breakdown is always on the pricing page.

### What happens when my agent restarts — does it lose its memory?

The session workspace and supported message history persist across ordinary pauses and restarts. If a machine is replaced, AgentSky restores the latest durable recovery point; this is managed recovery, not a zero-loss guarantee.

### How is my data handled?

AgentSky keeps the state needed to run and recover an agent. Data retention controls and any Zero Data Retention arrangement must be agreed and verified before they are enabled; they are not a self-serve product setting today.

### How is this different from a model API?

A model API returns model output; you still build and operate the agent around it. AgentSky serves the complete agent — harness, model, sandbox, tools, state, channels, recovery — behind one API, so what comes back is finished work, not tokens.

## Developer docs

Every docs page is also served as markdown at the listed .md URL; the
index is at https://agentsky.dev/docs/llms.txt and the full content at https://agentsky.dev/docs/llms-full.txt.
The machine-readable OpenAPI spec is at https://agentsky.dev/api/v1/openapi.json.

- [Developer quickstart](https://agentsky.dev/docs/index.md): Launch a long-lived agent from your terminal in two commands, then talk to it over a streaming API.
- [API reference](https://agentsky.dev/docs/api.md): Every /api/v1 endpoint with schemas, the error contract, and the stream event catalog.
- [Capabilities reference](https://agentsky.dev/docs/capabilities.md): Every built-in capability action agents can call — inputs, outputs, limits, and pricing units for web, image, video, and voice.
- [CLI reference — sky](https://agentsky.dev/docs/cli.md): Every sky command: auth, launch/deploy, agent, universe, and secrets.
- [agent.toml reference](https://agentsky.dev/docs/agent-toml.md): The declarative agent file — everything the create wizard can set, in a fly.toml-style file.

## Guides

- [Run your Buzz agent in the cloud](https://agentsky.dev/guides/buzz): point buzz-acp at an always-on cloud agent
- [Your agents in Telegram](https://agentsky.dev/guides/telegram): one bot for all your agents
- [Your agents in Slack](https://agentsky.dev/guides/slack): a channel per agent, teammates included

## Policies

- [Terms of Service](https://agentsky.dev/terms): the terms that govern use of AgentSky