raaaas

Parallel AI task orchestrator. Paste a task list into the desktop panel; Hermes phases it into dependency-ordered groups through the agent_batch_plan / agent_batch_phase tools (or you group it yourself in one click), then each phase dispatches one independent coding agent per task to a GitHub Actions workflow in a repo you name — own branch, own PR — and the panel tracks the runs and the PRs. GITHUB_TOKEN is the only required credential; AGENT_BATCH_REPO, AGENT_BATCH_WORKFLOW, AGENT_BATCH_MODEL and AGENT_BATCH_STATE are optional (the last is a directory override for the plan file, not a secret) and each can be set per plan in the panel instead. Disclosure — Hermes can read and rewrite the stored phase plan through the agent_batch_plan / agent_batch_phase tools, but dispatch is always a confirmed click in the panel. Each dispatch fires a GitHub workflow_dispatch with your GITHUB_TOKEN (Actions + pull-requests write) on the repo you name. The workflow you copy into that repo installs third-party coding-agent CLIs (cline by default; the others are unpinned npm packages) and runs them unattended, with auto-approve / skip-permissions, repo write access and your provider secrets. They push real branches and open real PRs, and your task text and repository code go to that runner's model vendor. LLM cost belongs to the runner: cline spends Cline Credits or a BYOK key, opencode free models need its zen relay, kilo needs an active Kilo Go subscription (keyless CI runs fail), and claude-code / codex / dsh need their own keys.