Harness Router · House #25,027

What it is

Calls itself “decision layer”.

• Calls itself: decision layer

“…decision layer”

Where it puts itself

Its one line of difference: “Jev and Sonnet 5 both reach 67.8% mean accuracy in TypeSafe AI’s published workflow evals. Jev’s reported cost is $0.0004 per case , compared with $0.1174 for Sonnet 5 — a 293.5× cost ratio”

“Jev and Sonnet 5 both reach 67.8% mean accuracy in TypeSafe AI’s published workflow evals. Jev’s reported cost is $0.0004 per case , compared with $0.1174 for Sonnet 5 — a 293.5× cost ratio”

What it costs

No free plan.

• Free plan: no

“MIT open source”

Who it sells to

Sells to developers.

• Sells to: developers

“Codex gets a native stdio server”

How it runs

A command-line tool. Runs locally and works offline. Offers an API. Open source. AI is the core of the product. You can bring your own API key. Built with GitHub Pages and Vercel.

• Kind: cli
• Runs where: on the device, offline
• API: yes
• Open source: yes
• AI role: core to the product
• Bring your own key: yes
• Built with: GitHub Pages · Vercel

“…uv tool install”

“Keep it isolated and globally available through uv tool .”

What it works with

Integrates with Codex and OpenRouter. Names Jev, MCTS and 3 more on its pages.

• Integrates with: Codex · OpenRouter
• Names: Jev · MCTS · Protocol Lattice · +2

“Use it as a Codex hook for automatic routing on every tool call, or as a skill when you want explicit, agent-invoked decisions.”

“It keeps obvious choices fast, uses Jev for ambiguity, and applies MCTS when multi-step consequences matter.”

What it offers as proof

Live, no waitlist. Claims “4,096 sims”. Publishes docs.

• Number claimed: 4,096 sims

“MIT open source”

“4,096 sims · depth 3 benchmark profile”