# AgentHydra: full brief Last updated: 2026-08-23 · Current version: v0.36.0 · License: MIT (free, open source) Home: https://agenthydra.lunarwerx.com/ · Source: https://github.com/LunarWerxs/AgentHydra ## What it is AgentHydra is a local dashboard that puts every Claude Code, Codex and OpenCode session on a machine into one list, manages isolated Claude and Codex instances, and queues `claude` runs to fire now or on a schedule. It is not a Claude client and does not replace one: it reads the local files those tools already write (Claude transcripts, Codex rollout JSONL, the OpenCode SQLite database) and serves a browser dashboard at `http://localhost:7787`, plus the same API over MCP stdio for agents. ## Who it is for Anyone running more than one AI coding CLI, or more than one account of the same CLI, on one machine: several Claude Desktop/CLI logins, a personal and work Codex account, or Claude Code alongside Codex and OpenCode, with no single place that shows all of that history at once. ## Key facts - Reads and lists Claude Code, Codex and OpenCode sessions together, newest first, badged by tool, searchable and filterable across all three. Also read-only lists the shared "foreign" session format used by Cursor, Windsurf, Zed and Copilot CLI where present. - Only Claude sessions can be messaged back; Codex/OpenCode are read-only live-tail, because only Claude ships a CLI that can be handed a prompt. - Manages isolated Claude Desktop/CLI instances (account, plan, weekly quota, live process/memory/uptime) and isolated Codex Desktop/CLI instances, each with its own `CODEX_HOME`, so a work and a personal login run side by side. - Queue and schedule `claude` runs with their own prompt, working directory, model, effort, permission mode and account; queued runs carry a start time and reattach after a restart. - Sessions stopped by a 5-hour rate limit can auto-resume once the window resets, gated on weekly usage, off by default. - Whole API over MCP stdio (sessions, queue, timing, instances, codex, usage tool groups) for Claude Code, Claude Desktop or Cursor. Agents can call `check_usage` before a heavy fan-out. - Quick instance mode (`AgentHydra.exe --instances`) skips the scanner/database/scheduler for a small start/focus/stop window on its own port. - Not entirely private: a periodic, opt-out update check against studio.connections.icu sends app version, a coarse OS tag, and a random per-install id (never an IP, hostname, username, file path, account or email). `AGENTHYDRA_NO_PING=1` routes the check to GitHub's API instead, with no install id. The optional ChatGPT handoff is the other thing that leaves the machine. - Built with Bun + Hono, Vue 3 + Vite, shadcn-vue, Tailwind v4, bun:sqlite, MCP stdio. ## Install - Windows: run the versioned `.exe` from Releases directly, no console window. - Linux/macOS: one-executable archives on Releases. No Bun needed. - From source: clone, `bun install`, `bun run build`, `bun run start` (needs bun.sh). - Dashboard binds to `http://localhost:7787` on loopback only either way. - Try without risk: `AGENTHYDRA_FAKE=1` stands in for the `claude` CLI so dispatch touches no real quota or repos. ## Pricing Free and open source, MIT license. No account, no paid tier. See `/pricing.md` for the machine-readable breakdown, including costs the user still carries themselves (their own Claude or OpenAI subscription/API key for dispatch). ## How it compares - **Separate terminals** (the default): one window per session, switched by hand, history scattered across whatever each CLI wrote. AgentHydra's purpose is one list instead. - **Crystal, now Nimbalyst** (stravu): an Electron app that launches new coding-agent sessions in parallel git worktrees for comparing approaches, covering Claude Code and Codex with OpenCode and Copilot in alpha. AgentHydra reads existing sessions instead of spinning up new worktrees. - **Conductor** (Melty Labs): a Mac app running a team of parallel Claude Code, Codex, Cursor and OpenCode agents, each in an isolated worktree, billed to the user's own subscription. AgentHydra runs on Windows/macOS/Linux as a dashboard over existing sessions. - **ccmanager** (kbwo, npm): a tmux-free terminal UI for Claude Code, Codex, Gemini and other CLI sessions across worktrees. AgentHydra is a browser dashboard plus MCP server, and also covers instance identity, a queue/scheduler, and quota checks. ## Limitations - Tray launcher is Windows only. macOS/Linux builds exist but are not yet verified. - Windows instance management needs the classic Claude Desktop installer, not the MSIX package. - Dispatch needs the user's own `claude` CLI. Codex/OpenCode support is optional, appearing automatically if their stores exist. ## FAQ Q: Is AgentHydra free? A: Yes, MIT-licensed, no account, no cloud service. Q: What tools does it support? A: Claude Code, Codex and OpenCode, side by side. Q: Can I reply to Codex/OpenCode sessions? A: No, read-only; only Claude ships a CLI to reply to. Q: Is it private? A: Mostly; a small opt-out update ping and the optional ChatGPT handoff are the only things that leave the machine on their own. Q: What can my agents do over MCP? A: Everything the dashboard does: sessions, queue, scheduler, instances, quota, via `check_usage` and the rest. Q: Does it work offline? A: Yes for reading/messaging; dispatch needs whatever Claude needs. Q: How does it differ from Crystal/Conductor/ccmanager? A: They launch new parallel sessions in worktrees; AgentHydra reads sessions already on the machine and adds a queue/scheduler/MCP layer. Q: Do I need Bun? A: Only to run from source; packaged builds need nothing else. ## Links - Home: https://agenthydra.lunarwerx.com/ - Source: https://github.com/LunarWerxs/AgentHydra - Download: https://github.com/LunarWerxs/AgentHydra/releases - Reference/MCP docs: https://github.com/LunarWerxs/AgentHydra/blob/main/docs/REFERENCE.md - Changelog: https://github.com/LunarWerxs/AgentHydra/blob/main/CHANGELOG.md - Pricing: https://agenthydra.lunarwerx.com/pricing.md - License: MIT, see the LICENSE file in the source repo - Parent studio: https://lunarwerx.com/