A hosting service that runs coding agents such as Claude Code and Codex in the cloud, with no servers to provision and nothing to set up. It positions itself as “the OpenRouter for agents”: a single API key lets you call several agent programs (harnesses) and several models interchangeably. Each agent runs on a persistent machine, so its conversation history and working state carry over as you talk to it from Slack, Telegram, WhatsApp, iMessage, the web, the CLI, or the API. While an agent sits idle, its machine suspends automatically and nothing is billed for that time. The service is based in the United States and took the number one Product of the Day spot on Product Hunt.
Key Features
- Mix and match harness and model: At launch you pick a harness from Claude Code / Codex / Hermes / OpenClaw / pi / DeepSeek Harness / Kimi Code / opencode, and a model from Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Grok and others. Either can be swapped later without losing the agent’s history
- One API call: Name the
"agent"and the"model"in your request and a cloud agent starts. Crashes, reconnects, and state persistence are handled by the service, so there is no container or session plumbing to build yourself - A resident agent across channels: Web, API, and CLI, plus Slack, Telegram, Discord, WhatsApp, iMessage, and the A2A protocol. Whichever channel you use, the agent shares the same history, the same tools, and the same state
- Move a local agent to the cloud as-is: A single
sky clonecommand copies the instructions, model, and MCP server configuration of the agent you already run locally. It keeps working while your laptop is closed - Bring your own subscription: Connect Claude Pro / Max and model usage for Claude Code agents is billed at $0; connect ChatGPT Plus / Pro and the same applies to Codex agents. You avoid paying twice for the same models
- Built-in tools and connectors: Web search, browser control, SEO data, image generation, video generation, transcription and more, plus over 2,000 app connectors including Gmail, Notion, Slack, GitHub, Google Drive, and Figma ─ all enabled as parameters on the same agent call
- Agent Arena (benchmarks): Elo scores derived from head-to-head runs on real tasks let you compare harness-and-model combinations on quality, speed, and cost
Pricing
There are no fixed monthly plans and no per-seat charges. You spend prepaid credits against two meters: compute (machine time) and model tokens.
| What is billed | Unit | Approximate rate |
|---|---|---|
| Compute (default 2 vCPU / 4 GB) | Per second, only while awake | $0.075 per hour / $1.81 per 24 hours / $54.33 per month if always on |
| Compute (smallest 1 vCPU / 512 MB) | Per second, only while awake | $0.011 per hour / $0.27 per 24 hours / $8.09 per month if always on |
| Suspended or parked agents | ─ | $0 (idle agents suspend automatically) |
| Storage | 1 GB persistent disk per agent | $0 |
| Model usage | Per 1M input / output tokens | e.g. Claude Sonnet 5 at $2.00 in / $10.00 out; DeepSeek V4 Flash at $0.44 in / $1.32 out |
| Model usage with a connected subscription | ─ | $0 for eligible agents when Claude Pro / Max or ChatGPT Plus / Pro is connected |
| Paid tools and connectors | Per use | e.g. neural search $0.01 per run, browser automation $0.03 per minute, connector call $0.029 per call |
No credit card is required to create an account, so you can start for free. Buying credits by card adds the payment processing fee (2.9% + $0.30), itemized at checkout.
Pricing is as of August 2026. Please check the official site for the latest rates.
Pros & Cons
✅ Pros
- No servers to build, monitor, or restart ─ you get a resident environment for coding agents immediately
- Harness and model can be replaced after the fact, which reduces lock-in to any one vendor
- Usage-based billing with no charge while idle keeps costs low for agents you run only occasionally
- An existing Claude or ChatGPT subscription can be reused, avoiding a second bill for the same models
- You can hand work to the agent from a chat app, without sitting in front of a terminal
⚠️ Cons
- Billing spans three meters (compute, models, tools), which makes the monthly total hard to predict in advance
- For always-on workloads it can cost more than renting a small VPS yourself
- Your code and credentials run on someone else’s cloud, so confidentiality needs checking for sensitive work
- With harnesses, models, channels, and connectors all to choose from, the first setup takes some deliberation
- It is a relatively new service, and long-term operational track record is still accumulating
Comparison with Similar Services
| Criteria | AgentSky | E2B | Daytona | OpenRouter |
|---|---|---|---|---|
| Main role | Agent hosting | Sandboxed execution for AI | Development workspace management | Model API aggregation and routing |
| What it provides | A harness plus a model | A sandbox for running code | A cloud development environment | Model endpoints |
| Harness included | Yes (Claude Code, Codex and others) | No (you implement it) | No (you supply it) | No |
| Chat channel integration | Slack, Telegram, WhatsApp and more | None | None | None |
| Billing | Compute plus tokens, usage-based | Usage-based on execution time | Usage-based on workspace uptime | Usage-based on tokens |
AgentSky’s position is that it bundles the harness, the execution environment, the channels, and the billing into one product ─ a wider scope than services that supply only the runtime or only the models.
Who Is It For
- Developers who want Claude Code or Codex to keep running away from their own machine
- Individuals and small teams who would rather not operate their own agent servers or container platform
- People who want to compare harnesses and models and pick the best combination per project
- Teams that want to try handing work to an agent from Slack or Telegram
- Product builders who want agents running behind their own service without owning the infrastructure
Summary
AgentSky takes over the question of where a coding agent lives. A design that lets you recombine harness and model, usage-based billing that charges nothing while idle, and control from chat apps together make it practical to run agents that are not tied to a local machine. Because the bill is the sum of several meters, though, the safe path is to start with a small machine size and an inexpensive model, watch actual consumption, and only then move to full operation. Whether you can bring your existing Claude or ChatGPT subscription makes a large difference to cost, so it is worth checking first.