Ctrl+Enter run · Ctrl+Shift+Enter kickstart
Pipeline
Template variables: {{task}} {{prev}} {{s1}}, {{s2}}… (each stage's output) · fan-out only: {{item}} {{index}} {{count}}
☁ Claude Code ultracode + Codex
Sends your task to the real Claude Code with ultracode on (Anthropic API or AWS Bedrock), OpenAI Codex and Google Gemini CLI, running in GitHub Actions. They work in parallel, then Claude merges every answer. Results arrive as comments on a GitHub issue, usually within 5–20 minutes. Runs are billed to the API keys saved as repository secrets, and only the repo owner can start one.
Or run it on your own machine
Typing the ultracode: keyword yourself inside Claude Code also works:
Claude Code (type in a session, v2.1.203+)
Codex CLI
Final output
Run one prompt through Claude, GPT, Gemini and AWS together.
- Add keys. Paste any of your API keys (Claude, GPT, Gemini, AWS Bedrock, or Any API such as OpenRouter) in Settings. Connect KakaoTalk there too. They stay in this browser and go only to the vendors' APIs. No keys yet? Leave Demo on to see the flow.
- Describe the task. Start it with
ultracode:if you want everything thrown at it. - Press ⚡ Kickstart. Relay picks the cheapest orchestration that fits and runs it. Every agent streams live, each stage feeds the next, and you see tokens and estimated cost as they happen.
Soloone agent
Compareboth, side by side
Relaydraft → cross-vendor attack → revise
Debateanswer → cross-examine → judge
Fan-outplan → workers → refute → synthesize
Councilevery connected AI answers, a judge merges
Ultraboth vendors at every step