# OptBench for coding agents

OptBench finds the LLM calls in your product, turns each into a use case with metrics and test cases, and benchmarks open and closed models on them. Your coding agent does all of it over MCP: you don't need the dashboard to start.

## Set up in one step

Add the MCP server to your agent, then sign in when it asks (your browser opens, no key to paste):

```bash
claude mcp add --transport http optbench https://app.optbench.com/mcp
```

In Claude Code, run `/mcp`, choose optbench, Authenticate. Cursor, Codex and other clients: [Clients](/docs/clients).

Then open your product repo and ask your agent:

```text
Scan this repo with OptBench.
```

The agent finds the LLM calls, proposes use cases with metrics and test cases, and shows you a cost estimate. Nothing that costs money runs until you agree.

## Pages

- [Clients](/docs/clients): setup for Claude Code, Cursor, Codex, VS Code and any MCP client
- [MCP server](/docs/mcp-server): the endpoint, OAuth or a workspace key, what an agent may do
- [Tools](/docs/tools): every tool, what it does, and whether it reads, writes, deletes or spends
- [Example prompts](/docs/prompts): what to ask your agent

Workflows:

- [Benchmark before launch](/docs/workflow-before-launch): scan, cases, pilot, run, and when a verdict is a Decision
- [Connect traces](/docs/workflow-connect-traces): OpenTelemetry, PostHog and Langfuse traces as test cases
- [Write gold](/docs/workflow-gold): the checks and answers a run is scored against
- [Check before you merge](/docs/workflow-drift): after a code change, which use cases no longer match the code
- [Monitor in production](/docs/workflow-production): what exists today and what is coming
- [Do it with your agent](/docs/with-your-agent): each screen, the tools that do the same, one prompt

Every page is also plain Markdown for agents: add `.md` to the URL (`/docs/clients.md`). The index for LLMs is [/llms.txt](/llms.txt).