shaxs

Benchboard — AI skill for Claude Code

AI community

Local-first benchmark for AI coding agents: one command runs an agent (OpenCode, Claude Code or any CLI) on a seeded React issue board, times it, counts tokens and scores it (80 automated + 20 human U.

How to install Benchboard

This entry records only its repository, not the path inside it, so there is no exact command to give. Open shaxs/benchboard and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Benchboard does

Alternatives in AI

Empryo — Empryo issue tracker + SoulForge (v2) 1.1k ★
Human In The Loop MCP Server — A powerful MCP Server that enables AI assistants like Claude to interact with humans through intuitive GUI dia 159 ★
Trekker — Simple opinionated issue tracker for AI Agents 74 ★

README

BenchBoard

BenchBoard is a local-first benchmark harness for AI coding agents, with a focus on local LLMs. A coding agent receives an existing React/TypeScript issue-board application (with seeded defects) as the fixed challenge and works on a disposable exported copy of it; the result is evaluated locally with an executable automated scorer worth up to 80 points, plus an optional human UI review worth up to 20 more points.

BenchBoard is released under the [MIT License](LICENSE). Its calibration is still pending (status `CALIBRATION_PENDING`): scores are meant for comparing agents, models and harnesses on your own machine, not as a calibrated public leaderboard or a certification.

Install

Node 22 or newer. Clone into an ordinary folder (not inside your system temp folder, which the evaluator keeps separate from the repository), then install the three dependency trees and Chromium:

git clone https://github.com/shaxs/benchboard.git
cd benchboard
npm ci
cd starter && npm ci && cd ..
cd reference && npm ci && cd ..
npx playwright install chromium
npm test

Then follow [docs/RUN.md](docs/RUN.md) to run an agent.

Start here

docs/RUN.md — run an agent (OpenCode, Claude Code or any CLI) end to end with one command, then score its UI on a local page
docs/QUICKSTART.md — a first walkthrough with generic placeholders
docs/MANUAL_WORKFLOW.md — the manual lifecycle end to end
docs/CLI.md — commands, --long-flag usage, exit codes and error categories (usage/state/integrity/infrastructure/cancellation)

What lives where

A run is a **container**: it holds a participant workspace and private run metadata, and `--run` names that container, not the workspace. The initial participant workload is the verified 26-file starter export. Protected challenge material (`spec/`, `starter/`, `fixtures/`, `reference/`, `runner/`, `evaluator/`) remains read-only in this owner che