Skip to main content
Bench connects your AI system, its prompts and your business context to repeatable evaluations. The quickest way in is the Set up using a coding agent prompt in onboarding or Getting started: one prompt for Claude Code or Codex connects the Bench MCP, installs the Bench SDK, sends one real request so Bench recognizes the system, registers every prompt, records what the system is for, runs the quick scan and, if you approve the spend, starts the first bench. Bench draws the system from that first trace: the runtime, each agent, the model calls and tools per agent, and hand-offs between agents. Connecting a GitHub repository is optional; use it later for background rescans and pull requests.
Use the setup instructions shown in your Bench account. Manual setup (SDK, GitHub, MCP) is available under Manual setup.

Set up with a coding agent

One prompt connects MCP, the SDK, prompts and understanding.

Run your first bench

Send one request, register model calls and evaluate.

Define success

Record purpose, rules, examples and system structure.

Build a test library

Keep trusted cases and customize criteria.

​
What a bench does

Bench creates criteria and scenarios, measures the current prompt, then tests candidate prompt and model changes on that suite. You can inspect the cases, scores, explanations and comparison before accepting a recommendation. Before the first bench, the quick scan already lists prompts whose quality is not measured yet, tools and hand-offs not tested yet, and each model’s estimated cost next to a lower-priced model in the same family.