soup-ai 0.2.0

Soup is a spec-compliant Agent Skills router for LLMs with deterministic BM25 routing: define many small skills and inject only the relevant ones into each call.

Soup

Soup is a provider-agnostic Agent Skills router for LLMs.

You define a large number of Agent Skills (rules, best practices, instructions) — in code, as SKILL.md files or with remote sources. On every LLM call, Soup injects only the skills that are actually relevant to the request.

Why?

Monolithic prompts do not scale once your skill catalog spans multiple domains, tools, and workflows. Soup routes only the relevant Agent Skills for each request, so you do not ship unrelated instruction blocks on every call.

• Reducing tokens sent on each call.
• Improving context quality (only relevant instructions).
• Avoiding monolithic prompts.
• Making context modular, composable and reusable across projects, using the open Agent Skills format so your skills are portable.

Soup is completely independent from the LLM provider. It works with OpenAI, Anthropic, Gemini, Ollama, LiteLLM, LangChain, etc. — because it just transforms the prompt/messages you already have.

Install

pip install soup-ai        # or: uv add soup-ai

Requires Python 3.12+. Runtime dependencies are pydantic and pyyaml.

Quickstart

from soup import Soup

soup = Soup()

soup.register(
    name="react-ui",
    description="React UI implementation patterns. Use for React components, hooks, and layout tasks.",
    instructions="""
Use React 19 function components.
Prefer composition and colocated state.
Keep accessibility semantics in markup.
""",
    tags=["react", "ui", "components"],
)

soup.register(
    name="sql-postgres",
    description="Postgres SQL guidance. Use for query design and indexing tasks.",
    instructions="""
Use parameterized queries.
Add indexes for frequent filters.
Review query plans before optimization.
""",
    tags=["sql", "postgres", "database"],
)

# String prompt in, string prompt out:
prompt = soup.prepare("I need guidance on building a React dashboard card component with hooks.")

# ...or chat messages in, chat messages out:
messages = soup.prepare(
    [{"role": "user", "content": "I need guidance on building a React dashboard card component with hooks."}]
)

The integration is meant to be invisible:

response = client.responses.create(
    model="gpt-5",
    input=soup.prepare(messages),
)

Registering skills

register() is the single entry point. It dispatches on what you give it.

# 1. Define a skill in code (name, description and instructions are required):
soup.register(
    name="pdf-processing",
    description="Extract PDF text, fill forms, and merge files. Use for PDF tasks.",
    instructions="Use pypdf for extraction...",
)

# 2. Load a single local skill directory (a folder containing SKILL.md):
soup.register("./skills/pdf-processing")

# 3. Load a whole local collection (a folder of skill directories):
soup.register("./skills")

# 4. Load from a GitHub or GitLab repo over raw HTTP (no git required):
soup.register("https://github.com/vercel-labs/skills", ref="main")
soup.register("https://gitlab.com/group/project", ref="main")

Soup's optional routing behavior stays simple. For a single skill, pass overrides as keyword arguments; for a collection or repo, pass per-skill options:

soup.register("./skills/pdf-processing", dependencies=["files"], version="1.0")

soup.register(
    "./skills",
    options={
        "pdf-processing": {"dependencies": ["files"], "priority": 10},
        "data-analysis": {"version": "2.1"},
    },
)

Core concepts

Skill

A Skill is the atomic unit of context and follows the Agent Skills spec.

Field	Group	Type	Purpose
name	required	str	Unique identifier (spec naming rules; matches dir name).
description	required	str	What it does and when to use it; primary routing signal.
instructions	required	str	The context injected; the SKILL.md Markdown body.
license	spec	str | None	License name or bundled file reference.
compatibility	spec	str | None	Environment requirements (max 500 chars).
allowed_tools	spec	list[str]	Pre-approved tools (allowed-tools, experimental).
metadata	spec	dict[str, str]	Arbitrary key-value mapping; stores Soup extensions.
tags	soup	list[str]	Keywords for tag-aware selection.
examples	soup	list[str]	Optional examples rendered after the instructions.
priority	soup	int	Tie-breaker; reserved for future compression.
dependencies	soup	list[str]	Skills to always include alongside this one.
extends	soup	list[str]	Parent skills this one specializes (composition).
version	soup	str | None	Version string for sharing/reuse.

SKILL.md files

A SKILL.md file is YAML frontmatter followed by a Markdown body. Soup's own routing extensions live under spec-compliant metadata (string values are canonical; comma-separated strings are parsed into lists):

---
name: pdf-processing
description: Extract PDF text, fill forms, and merge files. Use for PDF tasks.
license: MIT
metadata:
  version: "1.0"
  dependencies: "files"
  tags: "pdf, forms, documents"
  priority: "10"
---

Use pypdf for extraction. Validate inputs before merging...

The Markdown body becomes the skill's instructions. The name must follow the spec rules (lowercase letters, numbers and hyphens, 1–64 chars) and match its directory name.

Composition & versioning (extends)

Skills are composable and versionable, so you can build hierarchies and reuse rules across projects without duplication:

soup.register(name="frontend", description="Frontend basics. Use for UI work.",
              tags=["ui"], instructions="Accessibility first.")
soup.register(name="react", description="React rules. Use for React apps.",
              version="1.0", extends=["frontend"], tags=["react"], instructions="Use hooks.")
soup.register(name="nextjs", description="Next.js conventions. Use for Next.js.",
              version="2.1", extends=["react"], tags=["nextjs"], instructions="Use the app router.")

Selecting nextjs automatically pulls in react and frontend, rendered parent-first. dependencies work the same way (transitive inclusion) but express a companion relationship rather than specialization. Cycles are detected and broken, and missing references raise MissingDependencyError (configurable).

Selection strategies

By default, Soup uses a single built-in selector: BM25Strategy (no embeddings, no vector DB, no RAG).

If you need custom behavior, you can still add your own strategy:

from soup import SelectionStrategy

class RegexStrategy(SelectionStrategy):
    def select(self, query, skills):
        ...

soup.add_strategy(RegexStrategy())

For architecture and design details, see docs/architecture.md.

Examples

See examples/ for complete, runnable integrations with OpenAI, OpenRouter, Anthropic, LiteLLM and Ollama.

Development

uv venv && uv pip install -e ".[dev]"
ruff check .
mypy
basedpyright
pytest --cov=soup --cov-report=term-missing

Roadmap

V1 loads SKILL.md instructions and metadata only. Full Agent Skills support — exposing bundled scripts/, references/, and assets/ files to the LLM/runtime — is tracked as a follow-up.

Changelog

See CHANGELOG.md for release notes.

License

MIT.