经 AI Skill Hub 精选评估,VulcanBench 获评「强烈推荐」。这款AI工具在功能完整性、社区活跃度和易用性方面表现出色,AI 评分 8.0 分,适合有一定技术背景的用户使用。
VulcanBench 是一款基于 Python 开发的开源工具,专注于 AI、LLM、benchmark 等核心功能。作为 GitHub 开源项目,它拥有活跃的社区支持和持续的版本迭代,代码完全透明可审计,支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流,都能提供稳定可靠的解决方案。
VulcanBench 是一款基于 Python 开发的开源工具,专注于 AI、LLM、benchmark 等核心功能。作为 GitHub 开源项目,它拥有活跃的社区支持和持续的版本迭代,代码完全透明可审计,支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流,都能提供稳定可靠的解决方案。
# 方式一:pip 安装(推荐)
pip install vulcanbench
# 方式二:虚拟环境安装(推荐生产环境)
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install vulcanbench
# 方式三:从源码安装(获取最新功能)
git clone https://github.com/morganlinton/VulcanBench
cd VulcanBench
pip install -e .
# 验证安装
python -c "import vulcanbench; print('安装成功')"
# 命令行使用
vulcanbench --help
# 基本用法
vulcanbench input_file -o output_file
# Python 代码中调用
import vulcanbench
# 示例
result = vulcanbench.process("input")
print(result)
# vulcanbench 配置文件示例(config.yml) app: name: "vulcanbench" debug: false log_level: "INFO" # 运行时指定配置文件 vulcanbench --config config.yml # 或通过环境变量配置 export VULCANBENCH_API_KEY="your-key" export VULCANBENCH_OUTPUT_DIR="./output"
VulcanBench is an open-source harness for measuring how well LLM coding agents do real software engineering work. It runs a model against a task, keeps the task's hidden tests away from the agent, grades the result deterministically, and records everything: the full trace, the final patch, tokens, wall-clock, cost, and a reproducible replay command. It measures a model either through a raw API or through the product it ships in (Claude Code, Codex, Cursor, Grok Build, ZCode, Muse Code), at every reasoning-effort level the provider exposes.
Published results, model cards and methodology live at vulcanbench.com. The public record behind each report is in docs/results/.
git clone https://github.com/morganlinton/VulcanBench.git
cd VulcanBench
make setup
source .venv/bin/activate
vulcanbench --help
Prerequisites: Python 3.12 or newer, Git with Git LFS, and Docker Desktop for real runs. Node 20 or newer only if you want the dashboard.
Confirm the harness works end to end without spending anything. The mock model is deterministic and offline; --sandbox local is fine here because its commands are canned:
vulcanbench run --task hello-world --model mock:synthetic --sandbox local
Then build the sandbox image once and run a real model. Real runs execute model-written shell commands, so they default to a network-off Docker sandbox:
make sandbox-image # vulcanbench/sandbox:base (Python, Go, Node)
export ANTHROPIC_API_KEY=... # or OPENAI_API_KEY, XAI_API_KEY, ...
vulcanbench run --task hello-world --model anthropic:claude-opus-5
Each run prints its scores and cost and writes ./runs/<id>/{trace.jsonl, summary.json, replay.html, final.patch}. final.patch is a real git diff of the agent's edits and replay.html is a self-contained replay you can open in any browser. Traces, summaries and patches are secret-redacted and size-capped before they are written, so run artifacts are safe to publish. See docs/QUICKSTART.md for the longer walkthrough.
VulcanBench is an independent evaluation harness and is not affiliated with, sponsored by, or endorsed by any model provider. Model and product names identify the systems under test.
- Bring your own keys and subscriptions. Every call is made under your account and your agreement with that provider; VulcanBench never bundles or shares credentials. Staying within each provider's terms is your responsibility. - Outputs are for evaluation, not training. Run artifacts capture model outputs solely for scoring, inspection and reproducibility. Providers prohibit using their outputs to train competing models; do not use VulcanBench artifacts, or any published corpus of them, for that purpose. There is deliberately no feature that exports outputs as a training dataset.
This is not legal advice; consult the current provider terms for authoritative guidance.
| Provider | Spec | Key | Effort |
|---|---|---|---|
| OpenAI | openai:<model> | OPENAI_API_KEY | Responses API reasoning.effort when --effort is set; minimal and max supported |
| Anthropic | anthropic:<model> | ANTHROPIC_API_KEY | Messages API output_config.effort; extra-high maps to xhigh |
| xAI | xai:<model> | XAI_API_KEY | reasoning_effort; default is high and reasoning cannot be disabled, so sweeps should pass an explicit level |
| Meta | meta:<model> | META_MUSE_SPARK_API | minimal to xhigh map directly; an unset effort runs at a model-chosen level |
| DeepSeek | deepseek:<model> | DEEPSEEK_API_KEY | low, high, max; medium is recorded only |
| Alibaba Qwen | qwen:<model> | DASHSCOPE_API_KEY | low, medium, xhigh; high is recorded only |
| Moonshot Kimi | kimi:<model> | MOONSHOT_API_KEY | extra-high maps to max; others recorded only |
| Z.ai | zai:<model> | ZAI_API_KEY | recorded only |
| OpenRouter | openrouter:<vendor>/<model> | OPENROUTER_API_KEY | pinned to one upstream endpoint so a column stays one serving stack |
| Ollama | ollama:<model> | none | local inference through any OpenAI-compatible server; cost records as $0 and duration measures your hardware |
| Mock | mock:synthetic | none | deterministic, offline |
Optional. The backend serves ./runs as an API and the dashboard reads it:
pip install -e ".[backend]"
uvicorn backend.app:app --port 8000
cd dashboard && npm install && npm run dev # http://localhost:3000
Set DATABASE_URL (Postgres or SQLite) for a durable store; `docker compose up db provides Postgres and python scripts/ingest_runs.py` loads existing runs. See docs/DEPLOYMENT.md.
高质量的开源AI基准测试工具
AI Skill Hub 为第三方内容聚合平台,本页面信息基于公开数据整理,不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后,再部署至生产环境,并做好必要的安全评估。
✅ Apache 2.0 — 宽松开源协议,可商用,需保留版权声明和 NOTICE 文件,含专利授权条款。
AI Skill Hub 点评:VulcanBench 的核心功能完整,质量优秀。对于AI 技术爱好者来说,这是一个值得纳入个人工具库的选择。建议先在非生产环境试用,再逐步推广。
| 原始名称 | VulcanBench |
| 原始描述 | 开源AI工具:Open source, clear, transparent, real world llm benchmarks。⭐17 · Python |
| Topics | AILLMbenchmark |
| GitHub | https://github.com/morganlinton/VulcanBench |
| License | Apache-2.0 |
| 语言 | Python |
收录时间:2026-06-27 · 更新时间:2026-06-27 · License:Apache-2.0 · AI Skill Hub 不对第三方内容的准确性作法律背书。