Hydra
What does it take to supervise ten coding agents at once without losing track of any of them?
A desktop app that runs Claude Code and OpenAI Codex agents in parallel, with a manager agent that coordinates the others over MCP.
One window, many agents
Each agent runs in its own real PTY terminal, grouped by project. A chat view focuses on one agent; a grid view tiles a whole project and can broadcast one prompt to all of them. Workspace state survives restarts, and previous Claude sessions can be imported and resumed.
The interesting part is the manager: a special agent that runs its own HTTP/SSE MCP server, so another model can orchestrate the fleet through tools.
hydra_broadcast — Send the same prompt to every agent in one project.
Built to be operated
Headless runs
Queue background prompts without a terminal, with structured JSONL logs per run, search, filtering and session resume.
Budgets
Daily usage rollups per project and agent, soft token and USD budgets, and warnings at a configurable threshold.
Remote control
Scan a QR code on the desktop; a PWA on the phone joins the session with an explicit handshake and lists agents by project.
Observability
Rotating structured logs, renderer and main-process error capture, and an exportable diagnostics bundle.
One more section for engineers: architecture, hyper-parameters and design decisions.
Under the hood
Under the hood- Shell
- Electron 33, electron-vite + Vite 5
- UI
- React 18 + TypeScript, chat and grid views
- Terminals
- node-pty + xterm.js
- Orchestration
- Model Context Protocol SDK, HTTP/SSE server inside the app
- Validation
- Zod
- Quality
- Vitest + React Testing Library; CI lint, typecheck and tests on Ubuntu and Windows
- Packaging
- electron-builder; tag-triggered CI workflow for DMG, Windows installer, AppImage + .deb, and a Homebrew tap updater
What it doesn't do (yet)
- Usage and cost figures are parsed from CLI output: signals, not billing data.
- No public release has been cut yet, and builds are unsigned.