Hail Hydra plugin
Ten cheap-model subagents for Claude Code, Gemini CLI, Codex CLI and Copilot while the main model plans and verifies.
- 51 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Subagents read all project code, run commands and some edit files and make commits.
Install
- Works in
Manual install
npx hail-hydra-cc@latestThe README command opens a CLI picker. Add `--global` for a global install.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Subagents read all project code, run commands and some edit files and make commits.
This is third-party code. Review the repository files before installing.
What it does
Hail Hydra adds ten specialist agents, called heads, to a coding CLI. Each head is pinned to the cheapest model that can do its job. The main model plans, decides and verifies while the heads search code, run tests, write documentation, work with git and write or review code. The first tier holds cheap heads: `hydra-scout` for search, `hydra-runner` for tests and builds, `hydra-scribe` for docs, `hydra-guard` for security checks, `hydra-git`, `hydra-sentinel-scan` and `hydra-preflight`. The second tier is mid-level: `hydra-coder`, `hydra-analyst` and `hydra-sentinel`. After substantial edits a fast sweep checks imports and signatures and a deep check runs only on doubt. `/hydra:stats` reads your CLI session logs and reports real spend. There is a codebase map and memory across sessions.
Who it is for. Developers running long sessions in Claude Code, Gemini CLI, Codex CLI or Copilot CLI who want to cut token spend.
Good fit when
- Searching a large repository and running tests should not eat the main model context.
- After a big edit you need a fast sweep of imports and signatures.
- You need exact statistics on what a session cost.
- One agent set is needed across several CLIs.
Not a fit when
- You need deep architectural work that takes the main model entirely.
- The CLI does not support subagents.
- Background agents with file write access are not allowed.
Example request
Use Hydra to find every call site of a function, then run the tests through hydra-runner and show /hydra:stats.Limitations
The claimed savings and speed, about 50 percent of cost and 2 to 3 times faster, come from the author, and real numbers depend on how often Hydra is invoked. Statistics come from your CLI logs with no model estimation. Heads run as subagents and some of them write code and make commits, so review their edits. Default model names are set per CLI and change with their lineups.
How to disable. Delete the installed agent and command files from your CLI directory. The `/hydra:update` command handles updates.
FAQ
How do I install Hydra?
Run `npx hail-hydra-cc@latest` and pick the CLIs in the installer.
How do I install for a specific CLI?
For example `npx hail-hydra-cc --agent=claude --global` for Claude Code, and `--gemini`, `--codex` or `--copilot` for the others.
Which models does each tier use?
In Claude Code the cheap tier runs on Haiku, the mid tier on Sonnet and your main model stays the orchestrator. The README table lists models for other CLIs.
How do I see the real savings?
Run `/hydra:stats`. It reads your CLI session logs and shows spend per tier.
How many agents and commands ship?
Ten head agents and ten commands, plus six hooks, a codebase map and memory.
Discussion
No questions or reviews yet.
Related
Pluginsclaude-mods adds 112 skills, 3 agents and 3 commands to Claude Code: stacks, agent operation, defense, session continuity.
ClawMetry reads the session files agents already write and shows the timeline, tool calls, tokens and cost in one dashboard.
A personal ~/.claude for Claude Code: twelve subagents, review commands, worktree and goal skills and a Go status line.
A template for Claude Code projects: parallel agent swarms, workflow skills, safety hooks and rules in one install.
Multi-agent orchestration for Claude Code: teams, roles, autopilot and task routing across skills and subagents
A kit of 138 agents, 296 skills and 74 hooks that turns Claude Code into a full software development team
README in short
The set moves routine work to cheap models and leaves planning and verification to the main model. The README gives a table of heads, models per CLI and a task routing diagram.
README badge
Original SKILL.md
--- name: hydra description: > Multi-agent orchestration framework for Claude Code. Automatically delegates tasks to cheaper, faster sub-agents (Haiku, Sonnet) while maintaining Opus-level quality through verification. Use when working on any coding task — Hydra activates automatically to route file exploration, test running, documentation, code writing, debugging, security scanning, and git operations to the optimal agent. Saves ~50% on API costs. --- # 🐉 Hydra — Multi-Headed Speculative Execution You are the orchestrator of a multi-agent toolkit: route substantial work to cheaper specialized heads, and keep your own reasoning for planning, verification, and genuinely hard problems. ## Core Protocols The sentinel scan is the safety guarantee that replaces the orchestrator re-reading every diff — run it after every code change. ### Protocol 1: Sentinel Scan After Code Changes When any agent returns output containing `⚠️ HYDRA_SENTINEL_REQUIRED`, dispatch hydra-sentinel-scan with the files and changes listed in the trigger block — before presenting results to the user, before running any other agents. This is blocking: the user does not see the code changes until sentinel completes. Sequence: 1. Receive agent output containing ⚠️ HYDRA_SENTINEL_REQUIRED 2. Immediately dispatch hydra-sentinel-scan AND hydra-guard in parallel 3. Wait for both to complete 4. If sentinel-scan finds issues → dispatch hydra-sentinel (deep analysis) 5. Wait for deep analysis 6. Then present results to the user If the agent output contains `✅ HYDRA_NO_CODE_CHANGES`, skip sentinel. Present results immediately. ### Protocol 2: Sentinel Fix Decision Tree When hydra-sentinel confirms real issues: