Claw-Anything: Benchmark for Always-On Personal AI Assistants

Claw-Anything is the official benchmark and training set for evaluating always-on LLM agents, released by LiberCoders. The dataset assesses agents across three axes of real-world context: long-horizon event streams, interconnected services, and cross-device interaction. It includes a training set of 2,000 environments and was last updated in May 2026.

Use Cases

• Benchmarking agent performance on long-horizon tasks based on the described event streams.
• Training agents to handle interconnected digital services as mentioned in the description.
• Evaluating cross-device interaction capabilities across GUI and CLI interfaces.
• Researching the real-world context integration challenges for personal AI assistants.

Strengths

• Includes a training set of 2,000 distinct environments.
• Benchmarks performance across three defined axes of real-world context.
• Provides a quantitative baseline, with GPT-5.5 achieving only 34.5% pass@1.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full benchmark is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

LiberCoders

Freshness

Last updated 2026-05-26 06:39:06; freshness should be verified.