Skip to content
Ankit Jha edited this page May 29, 2026 · 2 revisions

Hydra — Wiki

Hydra is a multi-model AI orchestration system. Claude Code acts as the Brain (Tier 1 orchestrator) while cheaper, faster, or specialized models served via Antigravity CLI and local Ollama act as Heads (Tiers 2–10).

By routing tasks to the right model tier based on complexity, Hydra optimizes API costs, preserves Claude's context window, and accelerates developer velocity.


Quick Navigation

Page Description
Architecture System layers, orchestration protocol, and A2A handoff
Routing Protocol How enum keys map to tiers, fallback chains, pool exhaustion
Model Tiers All 10 tiers — models, executors, cost, and accuracy
Token Management Claude preservation thresholds and compaction rules
Dispatch Scripts Every script in dispatch/ documented
Registry Configuration All YAML config files in registry/
Getting Started Installation, CLI usage, and TUI guide

Core Concepts

The Brain / Head Split

Claude Code (Tier 1)          ← Brain: classifies, orchestrates, reviews
     │
  route.sh + decide.sh        ← Router: resolves enum → tier → model
     │
 ┌───┴──────────────────────┐
 │  Execution Heads (T2-T10) │
 │  Antigravity + Ollama     │
 └───────────────────────────┘
  • Brain — Claude Code. Does reasoning, routing decisions, and final review. Never call it for boilerplate.
  • Heads — Every other tier. Execute one delegated task, return raw output, escalate if blocked.

The Dispatch Command

# Preferred: dispatch by enum key
dispatch/route.sh --enum SIMPLE --prompt "write a User DTO in TypeScript"

# With file context
dispatch/route.sh --enum STANDARD --prompt "add pagination" --context src/users/controller.ts

# With A2A handoff from a prior agent
dispatch/route.sh --enum MODERATE --prompt "add auth" --a2a logs/last_handoff.json

# Check system status
dispatch/route.sh --status

The 10-Tier Enum

Enum Tier Model
CORE 1 Claude Code (direct)
EXPERT 2 Claude Opus 4.6 Thinking
VERY_HARD 3 Claude Sonnet 4.6 Thinking
COMPLEX 5 Gemini 3.1 Pro High
MODERATE 6 Gemini 3.1 Pro Low
STANDARD 7 Gemini 3.5 Flash High
SIMPLE 8 Gemini 3.5 Flash Medium
TRIVIAL 9 Gemini 3.5 Flash Low
GRUNT 10 Qwen2.5-Coder 7B (local)

Repository Links

Clone this wiki locally