WildClawBench
End-to-end real-world task evaluation for autonomous systems
All Skills in This Repository (9)
Pure Emerald Level Indicators
InternLMagent-browser
Automate headless browser navigation, interaction, and data extraction with Puppeteer.
InternLM03_task6
Extract and cross-check Slack messages to draft executive summaries.
InternLM03_task1
Automate meeting scheduling by integrating email and calendar APIs.
InternLMself-improvement
Log errors, corrections, and feature requests to structured markdown files.
InternLMvideo-frames
Extract frames or short clips from video files using ffmpeg.
InternLM03_task4
Extract and reconcile Slack messages to draft project status reports.
InternLM03_task3
Extract and analyze Slack messages to identify commitments, conflicts, risks, and deliverability issues.
InternLM03_task2
Extract action items, deadlines, and requests from Slack messages.
InternLM03_task5
Prioritize support messages, route issues to teams, and draft replies.
Frequently Asked Questions
FAQPage SchemaHow to install WildClawBench?▼
Run `npx skills add InternLM/WildClawBench --all -g -y` in your terminal to install all skills in this suite globally.
What does WildClawBench test?▼
It tests 60 original real-world tasks across six areas: productivity workflows, code intelligence, social interaction, search and retrieval, creative synthesis, and safety alignment. Tasks include clipping video highlights, negotiating meetings over email, and catching credential leaks.
Which models and harnesses does WildClawBench support?▼
The same 60-task suite runs under four harnesses: OpenClaw, Claude Code, Codex CLI, and Hermes Agent. Over 30 frontier models have been scored on the public leaderboard.
How does WildClawBench prevent cheating and data leakage?▼
Each task runs in its own Docker container with identical data and grading code. Ground truth and grading scripts are injected only after the run finishes, so they are never visible during execution.
Where can I find WildClawBench results and trajectories?▼
The interactive leaderboard is at internlm.github.io/WildClawBench, and full task data, Harbor-format packages, and complete run trajectories are published as Hugging Face datasets.
Related Repositories in Software Engineering
View All in Software Engineering→
openclawopenclaw
Run a personal AI assistant across your devices and chat apps
obrasuperpowers
Gives coding agents a disciplined workflow from idea to merged code
facebookreact
AI agent skills for building, testing, and porting React core