skillberry-ai/cap-evolve

View on GitHub ↗

Optimize any AI agent’s skills, tools/MCP, and prompts against your own evals.

38 ★11 forksPythonUpdated 18d ago

What you need to know

Host-agnostic harness that improves an AI agent's prompts, tools, and skills by learning from failed evaluation traces. Runs evaluate -> diagnose -> edit -> gate-accept with a val significance gate and a sealed test split. Zero runtime dependencies.

Install

pip install ./core

Usage

  • •Wire a small adapter (tasks/run_target/score) and run cap-evolve to optimize an agent's capabilities against your own eval.

Key features

  • ✓Optimizes system prompts, tool code, MCP surfaces, and skill packages
  • ✓Val-only significance gate with sealed test split
  • ✓Git-versioned candidates
  • ✓Live dashboard

Best for

Improving agent capability prompts/tools/skills with honest, measurable evaluation

Caveats

  • ⚠Beta (0.x)
  • ⚠Requires your own agent and eval harness
Documentation ↗

Reviewed 2026-08-11

Topics

agent-optimizationagent-skillsagentic-aiai-agentsclaudedspyevalsgepallmllm-evaluationmcpmodel-context-protocolprompt-engineeringprompt-optimizationpython
Stars
38★
Forks
11
Language
Python
License
Apache-2.0
Created
2026-06-15
Last push
2026-08-17