For developers
Pick any agent stack. Keep one simple API.
Choose the coding agent, model, tools, apps, and channels your product needs. Change the setup without rebuilding your integration.
One API · supported stacks · persistent sessions
Your product
AgentSky API# Configure a reusable agent
POST /api/v1/agents
{
"name": "code-reviewer",
"agentType": "codex",
"llm": "gpt-5.6-sol",
"prompt": "Review changes and explain risks."
}
# Start its computer and first task
POST /api/v1/sessions
{
"agent": "code-reviewer",
"initial_events": [{
"type": "user.message",
"parts": [{ "type": "text", "index": 0,
"text": "Review pull request #482" }]
}]
}
# Continue and watch the same session
POST /api/v1/sessions/{session_id}/messages
GET /api/v1/sessions/{session_id}/streamOne API
Choose what runs. Keep how you connect.
Every supported harness uses the same agent and session contract. Pick a compatible combination, send work, stream events, and continue through one API.
Harness
Claude Code, Codex, Hermes, or OpenClaw—all behind the same API contract.
Model
Choose a compatible Claude, GPT, Gemini, DeepSeek, Kimi, or GLM model for the work.
Tools + apps
Enable browser and search, then connect the apps and credentials the agent needs.
Channels
Use the same session from API, Web, Slack, Telegram, Discord, or the sky CLI.
Change the stack
Use another supported harness + model for new work. Your product keeps the same API contract.
Change the channel
Continue the same session from another connected channel with its workspace and history intact.
Customize the agent
Everything your agent needs, already behind the API.
Configure the worker once, then let AgentSky handle its computer, state, credentials, and connected surfaces. Change the configuration when the work changes.
Swap the worker
Choose a different supported harness and compatible model for the next job without changing your product integration.
Change its instructions
Update the prompt, reusable skills, tools, and per-session instructions as the job changes.
Connect its work
Add capabilities, packages, project data, and supported apps without building another runtime.
Keep access scoped
Credentials remain write-only and are injected only into the agent that needs them.
Connect once
Let the agent work anywhere.
Tools + apps
Enable browser and search in the agent spec. Connect supported apps such as GitHub, Gmail, Notion, and Slack in AgentSky.
Channels
Reach the same session from your API, Web, Slack, Telegram, Discord, or the sky CLI—without creating a separate agent for every surface.
Tools are configurable through the public agent API. Connector and channel setup is managed in AgentSky today.
The managed runtime
You build the product. AgentSky keeps the agent working.
A model endpoint returns tokens. AgentSky gives the harness a computer, persistent workspace, and the lifecycle needed to keep doing real work across turns.
A sandbox per session
The harness starts inside an isolated workspace with the files and compute it needs.
Persistent state
Files, harness state, memory, and history survive restarts and image updates on a durable volume.
Managed recovery
AgentSky replaces failed runtimes and restores their workspace from backups when needed.
Secure isolation
Each session runs in its own environment and receives only the credentials and capabilities you configure.
Events and streaming
One event contract across every harness.
Sending work returns immediately. Subscribe to the live session stream for output, or list persisted messages to recover after a disconnect.
agent.messageagent.reasoningagent.tool_useagent.tool_resultturn.status_idleLive SSE
Stream reasoning, tool activity, messages, status, and turn boundaries while the session runs.
Event history
Catch up from persisted user and agent messages with cursor pagination. Live-only tool events are not replayed.
Machine-readable contract
OpenAPI is generated from the same schemas used to validate every API request.
Scoped access
Use a revocable ast_ token with read, write, or admin scope. Token values are shown only when created.
Use the REST and SSE surface from any HTTP client today. The shipped developer client is the @agentsky/sky CLI; an official TypeScript or Python client SDK is not currently published.
Terminal quickstart
sky CLInpm i -g @agentsky/sky
sky auth login # authorize this machine via the browser
sky launch # scaffold agent.toml and create the agent
sky clone # or clone your local agent into the cloudAdd a working agent to your product.
Read the API contract, get a key, and start your first persistent cloud session.