AgentSky launches unified API for major code agents

The new platform AgentSky positions itself as “OpenRouter for agents”, a single call that can fire Claude Code, Codex, Hermes, pi, Kimi, OpenCode and any other major cloud agent. The developer picks an agent and model in the request; the system handles the rest—continuous compute, crash recovery, state persistence, GitHub integration, email forwarding or file upload. Registration is open without a credit card, and one API key unlocks all supported agents and models.

A company-released internal comparison shows a cost gap of two orders of magnitude: the same task on Claude Code cost $150, while running it on DeepSeek’s new agent harness cost $2. The figures are illustrative, not an official benchmark, but they highlight the economic logic of swapping agents when a cheaper model can finish the job for a fraction of the price. AgentSky lets developers change agent and model with a single parameter change, without rewriting the API-calling code.

The core technical component is a “persistent agent instance” that runs in the cloud. Crashes, disconnects and state restoration are handled automatically; developers do not manage machines, install dependencies or keep connections alive. Built-in connectors for channels such as Slack or Telegram and custom channels are passed as parameters in the agent call, eliminating separate configuration. The system also “pre-warms” agents so the first user request lands on a ready worker, avoiding cold-start latency.

The “Agent Hosting” feature copies a locally running agent—with the same instructions, model and MCP servers—to the cloud in one command. From then on it behaves as a regular cloud agent under the same API and continues running even when the laptop shuts down. Subscribing once to Claude Pro, Claude Max or ChatGPT lets agents run at zero model-usage cost, avoiding double billing for the same models.

Pricing splits into two parts: a per-minute charge for agent runtime and a per-token charge for model usage. This separation lets teams compare costs across different agents on the same model, or across models on the same agent, without mixing the two line items. For operations that fire thousands of tasks daily, that transparency is critical for budgeting.

The company’s “Agent Arena” logged 40,835 tasks between May and August 2026. Three early customers report sizable volumes: 150 thousand agents launched, 250 thousand tasks migrated from Claude Code to pi and open models with a single configuration change, and 50 thousand plus agents in production environments. The numbers come from the company and have not been independently verified, but they indicate real demand for an abstraction layer that decouples product logic from the specific agent executing it, including use cases where an AI CEO routes tasks.