cap-evolve plugin
cap-evolve edits an agent's prompts, tools and skills from failed traces and keeps an edit only if it wins on a held-out split
- 69 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- Evaluation runs call models and spend your account tokens
- The optimizer rewrites prompts, tool code and skills in the working copy and commits candidates
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/cap-evolveDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Evaluation runs call models and spend your account tokens
- The optimizer rewrites prompts, tool code and skills in the working copy and commits candidates
This is third-party code. Review the repository files before installing.
What it does
cap-evolve runs a loop: evaluate, diagnose the failures, propose an edit, check it on a held-out split and commit. It optimizes what the agent reads and leaves model weights alone. The target can be system prompts, executable tool code, MCP tool surfaces and whole skill packages, one or several jointly. An edit is accepted when the gain on the validation split exceeds a set threshold in standard errors. The test split is sealed and scored once. Every candidate is saved as its own git commit, and the dashboard shows cost, timing, diffs, lineage and a pass/fail matrix. It works through the `cap-evolve` CLI with `init`, `doctor`, `run`, `watch`, `diff` and `dashboard`, and Claude Code gets a plugin with slash commands. The README lists five algorithms, including GEPA and SkillOpt.
Who it is for. Authors of agents and skills who already have an evaluation set and want to improve prompts and tools from data.
Good fit when
- An agent has a set of test tasks while its system prompt is tweaked by eye
- A skill fails some tasks and you need to find the edit that fixes them without regressions
- MCP tool descriptions and code are optimized together with the prompt
- An improvement must pass a significance check before it lands
- You want to inspect the path of improvement: diff, cost and a tasks by iterations matrix
Not a fit when
- No evaluation set exists yet: the tool builds on an existing eval
- You need to fine-tune model weights: cap-evolve changes only the text and code around the model
- The job is a small one-off and the loop setup pays off only on repeated runs
Example request
Set up cap-evolve, run the keyless demo and show how the prompt score rose on the test splitLimitations
A ready evaluation of the agent comes first. Real runs call models and spend tokens, and `cap-evolve doctor` checks readiness before launch. Installation goes from a repository clone with `pip install ./core`, and the README mentions no PyPI package. Numbers in the README table come from the authors' own benchmarks, and your data will give different results. The Claude Code plugin loads with the `--plugin-dir` flag.
How to disable. Remove the virtual environment holding the `cap-evolve-core` package and the repository clone. The Claude Code plugin loads only by flag, so launch `claude` without it.
FAQ
How can I try cap-evolve without an API key?
Run `bash benchmarks/toy/examples/calc/run.sh` after installing the core. No model is called: a toy agent wins when the optimizer adds a marker to the prompt.
What does cap-evolve optimize?
The text and code around the model: system prompts, tool code, MCP descriptions and skill packages. Model weights stay the same.
How does cap-evolve check that an edit helps?
An edit is accepted when the validation gain exceeds a threshold in standard errors. The test split is sealed and scored once.
Where can I see the improvement history?
Each candidate is its own commit, and `cap-evolve diff --best` shows the edit from the seed to the winner. `cap-evolve dashboard` opens the browser panel.
Which algorithms are supported?
The README names five algorithms, including GEPA and SkillOpt, and `cap-evolve algorithms` prints the list.
Discussion
No questions or reviews yet.
Related
PluginsParallel self-evolution of skills and prompts: workers propose edits and an aggregator merges them into one versioned artifact
Runs a coding agent over tasks with known answers, studies failures and writes new skills and prompts, versioned in Git.
A plugin that cuts token use in long sessions: prunes stale tool output, picks a subagent model and narrows the skill list.
An agent evolution framework that edits prompts and skills with parallel workers and keeps only what passes a held-out check.
A GEPA plugin for Claude Code: the agent builds an evaluator and searches for a better prompt, code or config by score.
A Claude Code plugin with hooks and 26 skills: measures what the agent knows, blocks edits until code is read, keeps memory
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| cap-evolve (this card) | no | 69 | |
| GEPA optimize-anything | no | 7,415 | a plugin around the single GEPA algorithm, while cap-evolve holds several algorithms and a sealed test split |
| agent-eval | yes | 280 | scores an agent, while cap-evolve edits prompts and skills based on those scores |
README in short
The README opens with the idea of optimizing from evaluation failures and a list of strengths. A two-minute keyless example follows, then a CLI tour, the dashboard and a results table on several benchmarks. Below are run modes, the loop diagram and examples for different agent types.