How to load Skillware skills and connect them to language models. Each guide covers one provider adapter in SkillLoader.
SkillLoader.load_skill() accepts an absolute path to a skill directory, or a registry id such as compliance/tos_evaluator. When the id is not already a path on disk, the loader searches configured skill roots in resolution order.
Default (no config file):
- Roots listed in
SKILLWARE_SKILL_PATH(OS path separator between multiple roots) - A
skills/directory in the current working directory or its parents - Bundled skills installed with the
skillwarepackage (for example undersite-packages/skills/)
With config: copy .skillware.yaml.example to .skillware.yaml (or use global ~/.config/skillware/config.yaml) to persist project and external paths. Use the interactive menu (4 / paths) or edit YAML manually. Default order is project → external → bundled; the bundled registry is always included and remains the fallback when you have no local skills/ tree (typical after pip install skillware only). See CLI — config and skillware config show.
For pip-installed apps, bundled maintainer skills are the default; add private skills under ./skills/<category>/<name>/, config paths.external, or SKILLWARE_SKILL_PATH.
By default, SkillLoader.load_skill() validates manifest requirements before loading skill.py: unpinned deps must be importable; pinned specifiers (for example web3>=6.0.0) must match the installed version. See Install extras — Loader behavior.
Tip: Bundled registry skills are the default after
pip install skillware. Local or external skills with the same id shadow the bundled copy — seeskillware paths. Credential setup and provenance: API keys · Skill trust model.
To list locally available skills, inspect path resolution, show config, check load readiness, or run bundle tests from the terminal, see the CLI reference (skillware list, skillware paths, skillware config show, skillware doctor, skillware test).
| Provider | Adapter | Guide | Agent API key (typical) |
|---|---|---|---|
| Google Gemini | to_gemini_tool() |
gemini.md | GOOGLE_API_KEY (install skillware[gemini] for google-genai) |
| Anthropic Claude | to_claude_tool() |
claude.md | ANTHROPIC_API_KEY |
| OpenAI (ChatGPT) | to_openai_tool() |
openai.md | OPENAI_API_KEY |
| OpenAI-compatible hosts | to_openai_tool() |
openai_compatible.md | Host-specific key |
| DeepSeek | to_deepseek_tool() |
deepseek.md | DEEPSEEK_API_KEY |
| AWS Bedrock Converse | to_bedrock_tool() |
bedrock.md | IAM / AWS credentials; skillware[bedrock] |
| Azure OpenAI | to_openai_tool() |
azure_openai.md | Azure deployment credentials |
| Vertex AI (Gemini) | to_gemini_tool() |
vertex.md | GCP ADC / service account |
| Enterprise cloud (routing) | (see guide) | enterprise_cloud.md | Hosting + adapter choice |
| Ollama (prompt mode) | to_ollama_prompt() |
ollama.md | (local; no cloud key) |
| CLI | skillware list, skillware paths, skillware config show, skillware doctor, skillware test, skillware examples, skillware context, skillware chain |
cli.md | pytest in [dev] for test |
| Install extras | Category, skill, SDK, and meta pip install targets |
install_extras.md | See guide for [all], [agents], per-skill extras |
Skill-specific Usage Examples (sample prompts and execute payloads) live on each skill catalog page.
Shared patterns (load bundle, run execute, return tool results):
agent_loops.md. For multiple skills, first choose the host context, then use skill chaining (SkillContext, named chains:). Skill anatomy (Contract, Effect, Directive, Assurance, Interface). After load_skill, prefer bundle["class"]() to instantiate the skill; explicit bundle["module"].ClassName() also works. Runnable script inventory:
examples/README.md.
Contributors adding Usage Examples to skill catalog pages: skill_usage_template.md.
Skills that call external APIs during execution: API keys for skills.