Agent skills¶

Kaval.AI ships a set of skills for coding agents — short, task-shaped briefings that teach an agent working in your repository how this framework actually behaves. They matter because an agent’s priors come from other frameworks: without them it writes Jinja2 into a node prompt, inverts a similarity score, or treats a streamed failure as an HTTP error.

The skills are installed with the package, so there is nothing to download.

Skill	Load it when you are…
`kavalai`	starting out — install, extras, calling a model, structured output, and which of the others to read next
`kavalai-workflows`	writing or debugging a workflow graph, in YAML or with `WorkflowBuilder`
`kavalai-tools`	giving an agent tools: `@pythontool` functions, REST servers, MCP servers, and `allowed_tools`
`kavalai-serving`	serving, deploying, persisting or monitoring a workflow
`kavalai-rag`	indexing documents and retrieving them from a graph
`kavalai-eval`	writing evaluation cases, or testing a workflow without a model

kavalai-workflows and kavalai-serving carry reference files the agent reads only once it needs them — the exhaustive node tables, the load-time error messages, and every environment variable.

Installing them into a project¶

$ kavalai-skills install
installed  .claude/skills/kavalai
installed  .claude/skills/kavalai-workflows
...

--target chooses another directory (the default is .claude/skills), and a skill that is already there is kept rather than overwritten — reported as exists … (--force to overwrite) — so your own edits survive an upgrade unless you pass --force. Name one or more skills to install just those:

$ kavalai-skills install kavalai-workflows kavalai-tools --target .agents/skills

kavalai-skills list prints what is bundled, with each skill’s description.

The files also sit inside the installed package, at kavalai/.agents/skills/<name>/SKILL.md, which is where an agent that reads an installed distribution’s skills will find them without any install step.

Keeping them honest¶

A skill is documentation an agent reads instead of these pages, and it cannot tell when it has gone stale. tests/test_skills.py pins the parts a rename would silently invalidate — every node type, every node input type, every eval matcher, every key of EvalCase and EvalSuite, and every environment variable a skill names — against the code itself. Renaming any of them fails the suite until the skill is updated too.