Agent Builder
Launch your own agent. Live from day one.
Give it instructions, tools, and a model — in the UI or in code with sky launch. It goes live with its own API, its own channels, and hosting managed by AgentSky.
What it is
A UI and CLI for configuring, launching, and managing custom agents. Each agent is a named spec: harness, model, instructions, tools, channels.
Models
Any compatible harness + model pair from the AgentSky catalog — Claude Code, Codex, Hermes, and more. Switch the model without rebuilding the agent.
Cost
$0 while the agent is parked. Compute billed per second when a session is active. Model tokens at list price, or $0 with your own connected subscription.
What every agent gets
Live from day one.
Instructions and persona
Write a system prompt in the UI or TOML. Name it, shape its behavior, pin its memory defaults.
Tools and skills
Enable built-in skills (web search, browser, code exec) or connect external tool APIs from the catalog.
Channels
Connect Slack, Telegram, Discord, the web, or CLI. Every channel shares the same session state.
Connected apps
Link GitHub, Notion, Linear, and others. The agent reads and writes them without extra glue code from you.
Its own API
Every launched agent is addressable through the AgentSky session API — send work, stream events, continue sessions.
Hosting and recovery
AgentSky manages the cloud computer, state persistence, and restart recovery. You configure; it stays up.
In the UI
No code. No infrastructure.
Name the agent, write its instructions, pick a harness and model, and enable tools and channels. Hit Launch and it's live. Edit, update, and clone from the same dashboard.
Open the builder →In code — sky CLI
npm i -g @agentsky/sky
sky auth login # authorize via the browser
sky launch # scaffold agent.toml + create
sky clone # clone local agent to the cloudHow it works
Three steps to a live agent.
- 01
Configure in the UI or TOML
Name your agent, write its instructions, pick a harness and model, and enable tools and channels. No code or infrastructure required.
- 02
Launch with one command
Click Launch in the browser, or run sky launch to scaffold agent.toml and push it live from the terminal.
- 03
Reach it everywhere
The agent is immediately reachable via its API, connected channels, and the Playground. Share the link or wire up your product.
FAQ
Common questions.
What is the difference between the Builder and the Playground?
The Playground lets you compare any agent and model combination in a one-off session — no configuration saved. The Builder creates a named, persistent agent spec that remembers its instructions, tools, and channels across sessions.
Can I update an agent's instructions after launching it?
Yes — instructions can be updated at any time via the UI or the API (PUT /v1/agents/{id}/prompt). The updated instructions take effect on the agent's next session; existing sessions keep the prompt they started with.
How do I connect my agent to Slack?
From the agent's settings page, open the Channels tab, select Slack, and follow the OAuth install flow. The agent starts receiving mentions in Slack immediately; replies go back into the thread.
Can I use my own Claude or OpenAI subscription for model billing?
Yes. Run sky auth connect to link a Claude Pro/Max or OpenAI subscription. Eligible model usage costs $0 instead of per-token billing. Compute and connector costs remain separate.
Your agent, live in minutes.
Configure in the UI or code. No card required to get started.