Hydra

Local-first trust control plane that routes AI tasks to the cheapest model that clears your confidence bar.

Free· Free and open-source under the MIT license; no hosted tier or paid plan. You still pay whatever the underlying providers (Anthropic, OpenAI, OpenRouter, etc.) charge for tokens Hydra dispatches to them.AgentsMulti-provider: routes across Claude, GPT, Gemini Flash, OpenRouter-hosted models, and local Qwen via Ollama / LM Studio

Visit website →

Best for

Individual developers and small engineering teams who already juggle Claude, GPT, Gemini, and local Ollama/LM Studio models and want a scriptable, privacy-preserving router that keeps routine work off frontier APIs.

Skip if

Non-developers, teams that need a hosted control plane with SSO and org-wide dashboards, or single-provider shops that just want one model — Hydra's payoff comes from routing across many heads.

Hydra (CLI: hyctl) is an open-source, local-first orchestration layer for multi-model AI workflows. Instead of forcing every request through a single hosted router like OpenRouter, LiteLLM, or Portkey, Hydra discovers every 'head' already available on your machine — installed CLI agents (Claude Code, Codex, Gemini CLI), API keys in your environment, local Ollama or LM Studio servers, and OpenRouter-backed providers — and then dispatches each task to the cheapest head that can clear a configurable confidence threshold (default 0.95). It categorizes work across a ten-tier cost ladder and uses Sequential Probability Ratio Testing (SPRT) to stop querying as soon as the target confidence is met, so boilerplate rewrites don't burn frontier-model tokens and only genuinely hard tasks escalate to Opus- or GPT-class models. A 'blast radius' analyzer inspects your dependency graph and raises the confidence bar for changes that touch widely-used code, while a bundled local Qwen model acts as an offline failsafe so routing keeps working with zero network calls. An on-device 'accountability ledger' records which head handled what, giving teams an audit trail without shipping prompts or code to a hosted control plane. Typical workflows: point hyctl at your repo, run `hyctl probe` to enumerate heads, then use `hyctl dispatch` in place of your usual model CLI so refactors, tests, and doc-writing get routed locally while architecture questions escalate. Reported outcomes on the landing page include 58% of tasks staying fully local, ~73% median cost reduction versus always-frontier routing, and roughly a dozen heads discovered on a typical dev machine.

Editor's take

Hydra is the rare meta-router that respects your keys and your machine — decisions happen locally, the ledger stays local, and it slots into an existing multi-CLI setup instead of demanding you re-pipe everything through a hosted proxy. The SPRT + blast-radius framing is smart, but the headline savings numbers are self-reported and the project is still early, so treat it as a promising power-user tool rather than a settled default.

— The AI Tool Bible editorial team

Pros

✅ Genuinely local-first — routing decisions and the accountability ledger stay on your machine, unlike hosted meta-routers.
✅ Discovers heads you already have (Claude Code, Codex, Ollama, LM Studio, OpenRouter keys) instead of re-plumbing you through one vendor.
✅ SPRT-based confidence stopping avoids burning frontier tokens on tasks a cheaper model already answered well.
✅ Blast-radius heuristic ties confidence requirements to code impact, so trivial edits go cheap and load-bearing changes escalate.
✅ Ships with a local Qwen failsafe so dispatch keeps working offline or when an API is down.
✅ MIT-licensed and installable via brew, npm, pip, or a shell script — easy to adopt or fork.
✅ Provider-neutral by design; no lock-in and no need to hand a third party your API keys.

Cons

⚠️ CLI-only — no hosted UI, dashboard, or documented HTTP API for non-terminal workflows.
⚠️ Reported cost-savings numbers (73% median, 58% local) come from the vendor's own landing page and aren't independently benchmarked.
⚠️ Confidence scoring and blast-radius weighting are heuristics; miscalibration can silently route hard tasks to weak models.
⚠️ Value depends on already having multiple model backends installed and configured — thin benefit for single-provider users.
⚠️ Young project on a personal-namespace domain (uvansa.com / github.com/ankit373) with limited community track record versus LiteLLM or OpenRouter.
⚠️ No team/org features documented — the on-device ledger doesn't obviously roll up across multiple developers.

Use cases

multi-model CLI routingcost-optimized code generationoffline AI coding with local fallbackconfidence-gated task dispatchblast-radius-aware refactorson-device audit ledger for AI usageboilerplate work on cheap/local modelsescalation of hard tasks to frontier modelsvendor-neutral agent orchestration