Install extras

Skillware ships as one PyPI wheel. Every bundled registry skill is included on disk after pip install skillware. Optional extras add Python packages only — they do not download or hide skill bundles.

Use this guide to choose the smallest install that matches what you run. For loading skills after install, see Finding skills on disk.

Quick reference

Goal	Install command
Framework + all skills on disk (no optional runtime packages)	`pip install skillware`
Recommended: one bundled skill (always use this in docs)	`pip install "skillware[<category>_<skill>]"`
All skills in a category	`pip install "skillware[<category>]"`
Every bundled skill's runtime deps	`pip install "skillware[all]"`
Agent SDK adapters (Gemini, Claude, OpenAI)	`pip install "skillware[gemini]"` (or `[claude]`, `[openai]`, `[agents]`)
Clone-repo development + tests	`pip install -e ".[dev,all]"`
Development + agent SDK examples	`pip install -e ".[dev,all,agents]"`

Pip rule: pip install "skillware[extra]" always installs core + extra dependencies. There is no extra-only install without the skillware package.

Shell quoting: Use quotes on zsh and fish — e.g. pip install "skillware[finance_wallet_screening]".

Docs convention: Catalog pages, examples, and skill guides always recommend the per-skill extra (category_skill), even when it is empty today. When a skill gains new manifest requirements, run python scripts/sync_extras.py — docs stay unchanged.

Base install

pip install skillware

Includes:

Core framework (SkillLoader, CLI, discovery)
All bundled registry skills under site-packages/skills/
Core runtime dependencies: requests, pyyaml, python-dotenv, beautifulsoup4, packaging, jsonschema, rich

Does not include optional skill runtime packages (for example web3, fastembed, pymupdf, google-genai) or agent SDK packages (anthropic, openai). If you load a skill that needs them, SkillLoader raises an ImportError with suggested extras (see Loader behavior).

Which extra should I use?

flowchart TD
    A[What are you running?] --> B{Single skill}
    A --> C{Whole category}
    A --> D{Many skills / full registry}
    A --> E{Agent SDK examples}

    B --> B1["pip install skillware[category_skill]"]
    C --> C1["pip install skillware[category]"]
    D --> D1["pip install skillware[all]"]
    E --> E1["pip install skillware[gemini] or [claude] or [openai] or [agents]"]

    B1 --> F[Also need SDK? Add gemini/claude/openai/agents]
    C1 --> F
    D1 --> F

Single skill in production or documentation — always use the per-skill extra (pip install "skillware[category_skill]"), even when it adds no packages yet.
Exploring a domain — use the category extra.
CI, contributors, or multi-skill apps — use [all].
Runnable examples under examples/ — list each skill's extra plus an SDK extra ([gemini], [claude], …) when the provider is not local execute.

Skill extras are orthogonal to SDK extras. Example: office/pdf_form_filler needs [office_pdf_form_filler] for pymupdf and anthropic; a Gemini agent loop around it also needs [gemini].

Category extras

Union of non-core requirements from every skill in the category.

Extra	Skills	Packages installed
`compliance`	`compliance/mica_module`, `compliance/pii_masker`, `compliance/tos_evaluator`	`google-genai`
`creative`	`creative/bg_remover`	`rembg`, `pillow`, `onnxruntime`
`data_engineering`	`data_engineering/novelty_extractor`, `data_engineering/synthetic_generator`	`fastembed`, `numpy`
`defi`	`defi/evm_tx_handler`	`web3>=6.0.0`
`dev_tools`	`dev_tools/issue_resolver`	(none today)
`finance`	`finance/uk_companies_house_handler`, `finance/wallet_screening`	(none today)
`monitoring`	`monitoring/token_limiter`	(none today)
`office`	`office/pdf_form_filler`	`anthropic`, `pymupdf`
`optimization`	`optimization/prompt_rewriter`	(none today)
`security`	`security/prompt_injection_firewall`	(none today)
`wellness`	`wellness/mental_coach`	`google-genai`

pip install "skillware[defi]"

When a new category or skill lands, run python scripts/sync_extras.py after merging the skill manifest — category and skill rows appear in pyproject.toml automatically. Update this guide's tables when adding a category or skill with new packages.

Skill extras

One extra per bundled registry skill. Naming: {category}_{skill_name} (registry / becomes _).

Extra	Registry ID	Packages	Notes
`compliance_mica_module`	`compliance/mica_module`	`google-genai`
`compliance_pii_masker`	`compliance/pii_masker`	(none today)	Use this extra in docs and installs
`compliance_tos_evaluator`	`compliance/tos_evaluator`	(none today)	Use this extra in docs and installs
`creative_bg_remover`	`creative/bg_remover`	`rembg`, `pillow`, `onnxruntime`
`data_engineering_novelty_extractor`	`data_engineering/novelty_extractor`	`fastembed`, `numpy`
`data_engineering_synthetic_generator`	`data_engineering/synthetic_generator`	(none today)	Use this extra in docs and installs
`defi_evm_tx_handler`	`defi/evm_tx_handler`	`web3>=6.0.0`
`dev_tools_issue_resolver`	`dev_tools/issue_resolver`	(none today)	Use this extra in docs and installs
`finance_uk_companies_house_handler`	`finance/uk_companies_house_handler`	(none today)	Use this extra in docs and installs
`finance_wallet_screening`	`finance/wallet_screening`	(none today)	Use this extra in docs and installs
`monitoring_token_limiter`	`monitoring/token_limiter`	(none today)	Use this extra in docs and installs
`office_pdf_form_filler`	`office/pdf_form_filler`	`pymupdf`, `anthropic`
`optimization_prompt_rewriter`	`optimization/prompt_rewriter`	(none today)	Use this extra in docs and installs
`security_prompt_injection_firewall`	`security/prompt_injection_firewall`	(none today)	Offline-only; no runtime deps
`wellness_mental_coach`	`wellness/mental_coach`	`google-genai`

pip install "skillware[finance_wallet_screening]"

Empty extras ([]) are intentional — always use the per-skill extra in documentation and install commands so new manifest requirements do not require doc rewrites.

Meta extras

Extra	Purpose	Packages
`all`	Deduped union of all bundled skill runtime deps (non-core)	`anthropic`, `fastembed`, `google-genai`, `numpy`, `onnxruntime`, `pillow`, `pymupdf`, `rembg`, `web3>=6.0.0`
`agents`	Union of all agent SDK extras	`google-genai`, `anthropic`, `openai`
`dev`	Clone-repo lint and test tools	`pytest`, `pytest-mock`, `flake8`, `black`

[all] does not include SDK packages. For full local development matching skill bundle tests and agent examples:

pip install -e ".[dev,all,agents]"

Agent SDK extras

For SkillLoader.to_gemini_tool(), to_claude_tool(), to_openai_tool(), and provider examples — not required for execute() on skills that do not call that SDK internally.

Extra	Package	Guide
`gemini`	`google-genai`	gemini.md
`claude`	`anthropic`	claude.md
`openai`	`openai`	openai.md
`agents`	all three SDKs	agent_loops.md

pip install "skillware[gemini]"

Loader behavior

When SkillLoader.load_skill(..., check_requirements=True) (default) validates manifest requirements:

Unpinned entries (for example requests) — the importable module must exist.
Pinned entries (for example web3>=6.0.0) — the module must exist and the installed distribution version must satisfy the specifier.

On failure it raises ImportError with:

Missing packages and/or version mismatches (required spec vs installed)
Suggested pip install "skillware[<category>_<skill>]"
Category and [all] fallbacks

The loader does not install or upgrade packages — use pip extras or install the requirement strings yourself.

Use SkillLoader.load_skill(..., execute_module=False) for inspect-only loads (manifest, instructions, card, and optional requirement pre-flight) without executing skill.py. The CLI skillware doctor command uses this mode for the DEPS column before attempting a full import for LOAD.

Packaging smoke tests use check_requirements=False so a base wheel install can verify bundles without optional extras (TESTING.md).

Contributors

List runtime packages in the skill's manifest.yaml requirements (source of truth).
Run python scripts/sync_extras.py to regenerate category, skill, and [all] rows in pyproject.toml.
CI and tests/test_extras_sync.py fail if generated extras drift from manifests.

Core dependencies (requests, pyyaml, beautifulsoup4, …) should not be duplicated in extras — the sync script filters them automatically. Manifests may still list them for documentation; use bs4 or beautifulsoup4 interchangeably for Beautiful Soup.

Hand-maintained extras (dev, gemini, claude, openai, agents) live above the generated block in pyproject.toml and are not touched by the sync script.

See Packaging (PyPI and pip install) in CONTRIBUTING.md.

Removed legacy extras

Clean break in this release:

Removed	Replacement
`cli`	CLI ships on every install via `[project.scripts]`
`embeddings`	`data_engineering` or `data_engineering_novelty_extractor`
Old `[all]` (mixed SDK + skills)	`[all]` = skill runtime only; use `[agents]` for SDKs

Navigation Menu

Uh oh!

Install extras

Quick reference

Base install

Which extra should I use?

Category extras

Skill extras

Meta extras

Agent SDK extras

Loader behavior

Contributors

Removed legacy extras

See also

Uh oh!

FilesExpand file tree

install_extras.md

Latest commit

History

install_extras.md

File metadata and controls

Install extras

Quick reference

Base install

Which extra should I use?

Category extras

Skill extras

Meta extras

Agent SDK extras

Loader behavior

Contributors

Removed legacy extras

See also