{repo, branch, task}, get a PR URL back. Stateless. Streaming events for live UIs. Bring your own Anthropic, OpenAI, or open-weights key. Model tokens are billed by your provider, never marked up.POST /v1/tasks
{
"repo": "acme/api-server",
"branch": "main",
"task": "Add Redis-backed rate limiting (60 req/min per IP) to /v1/*",
"model": "claude-sonnet-5"
}
← 200 text/event-stream
event: status data: { "status": "cloning" }
event: tool_use data: { "name": "edit", "file": "middleware.ts" }
event: tool_use data: { "name": "edit", "file": "lib/limiter.ts" }
event: git data: { "action": "committed", "files": 2 }
event: git data: { "action": "pushed", "branch": "orchestra/redis-rate-limit" }
event: result data: { "pr_url": "https://github.com/acme/api-server/pull/318" }Trigger from CI, from a backend job, from a Slack bot, from your own product. Anywhere you can curl.
No session to maintain. Every request carries its own repo, token, and model key. Survives restarts, scales horizontally.
SSE for every tool call, edit, and commit. Render the agent's work live in your own UI. Same shape as Claude Code's native stream.
Pass your Anthropic / OpenAI / OpenRouter key on each request. We don't store it. We don't mark it up.
Same payload routes Sonnet, GPT, GLM, DeepSeek, Qwen. Pick per task: cheap on triage, premium on hard refactors.
Every response includes prompt tokens, completion tokens, and dollar cost so you can pass-through bill or budget per task.
The agent runs on your computer with the Claude, Codex or OpenRouter account you already have. Your code stays there. Model usage is billed by your provider; we never resell it.
Run coding agents on your own computer.
Take your desktop tasks with you, and work on them together.
Tell us where you'd integrate the API and what model you'd want to use. We'll send back an API key and a working example for your stack.