If you've used pydantic-ai, you already know the feeling: it's the first agent framework that feels like a regular Python library. Typed RunContext, a clean FunctionToolset protocol, model providers swapped behind one string. After a decade of frameworks that pretended to be Pythonic, this one actually is.

And then you sit down to wire up your agent, and realize: pydantic-ai will happily call any tool you give it — but the tools themselves are still on you. Want the agent to read a file? You write the sandbox. Run a SQL query? You write the read-only guard and the schema introspection. Search local documents? Text splitter, vector index, cosine math, persistence — all you.

After the third project where you wrote those by hand, the shape stops being interesting. This is them, written once:

pip install pydantic-ai-toolbox

from pydantic_ai import Agent
from pydantic_ai_toolbox import (
    FilesystemToolset, SQLToolset, PandasToolset, MemoryToolset, RAGToolset,
)

agent = Agent(
    "openai:gpt-4o-mini",
    toolsets=[
        FilesystemToolset(root="./workspace", read_only=False),
        SQLToolset(dsn="postgresql://user:pwd@localhost/app"),
        PandasToolset(),
        MemoryToolset(storage_path="./memory.json"),
        RAGToolset(embedder=my_embedder),
    ],
    system_prompt="You are a data assistant.",
)

print(agent.run_sync("Read README.md from the workspace and summarise it.").output)

That's the whole story. Five toolsets, one toolsets=[...], no new framework on top of pydantic-ai — each toolset is a thin FunctionToolset subclass, exactly what pydantic-ai expects.

Each toolset in isolation — runnable against a local/remote Ollama, no API keys:

• Filesystem — Example Create / Read / Append / Delete files
• SQL — Example INSERT, UPDATE, SELECT via SQLite
• Pandas — Example load a CSV, count rows by condition
• Memory — Example three-turn conversation with persisted facts
• RAG — Example retrieve-then-answer, override the model prior
• Quickstart — Example minimal one-tool smoke test

pip install pydantic-ai-toolbox

The base install gives you FilesystemToolset and MemoryToolset (stdlib only). The rest are opt-in so you only pull in what you use:

pip install "pydantic-ai-toolbox[sql]"      # + SQLAlchemy
pip install "pydantic-ai-toolbox[pandas]"   # + pandas + pyarrow
pip install "pydantic-ai-toolbox[rag]"      # + numpy
pip install "pydantic-ai-toolbox[all]"      # everything

Extras are independent — picking up one doesn't pull in the others. Details: docs/INSTALL.md.

Tiny snippets to taste each one:

# Filesystem — sandbox a workspace, then let the agent edit files
FilesystemToolset(root="./workspace", read_only=False)

# SQL — read-only Postgres
SQLToolset(dsn="postgresql://user:pwd@localhost/app")

# Pandas — start with an empty registry, agent loads CSVs as needed
PandasToolset()

# Memory — persisted scratchpad, 200-message cap
MemoryToolset(storage_path="./memory.json", max_messages=200)

# RAG — bring your own embedder
RAGToolset(embedder=lambda texts: [embed(t) for t in texts])

Runnable end-to-end scripts live in examples/ (see docs/EXAMPLES.md).

Before reaching for a toolset here, check whether pydantic-ai already ships the capability you need — most of the time it does:

This package fills the gaps that aren't on that list — local sandboxed filesystem access, generic SQL via SQLAlchemy, in-memory dataframe ops, self-hosted conversation memory, and local RAG without an external vector DB.

A toolset is a BaseToolset subclass whose public methods carry @tool:

from pydantic_ai_toolbox import BaseToolset, tool


class WeatherToolset(BaseToolset):
    """Look up current weather for a configurable provider."""

    def __init__(self, api_key: str, units: str = "metric") -> None:
        self.api_key = api_key
        self.units = units
        super().__init__()          # MUST be last — scans @tool methods

    @tool
    def current_temperature(self, city: str) -> float:
        """Return the current temperature for `city` in the configured units."""
        ...

Full rules, schema-mapping table, and the contribution checklist: docs/CUSTOM.md, AGENTS.md.

pip install git+https://github.com/wachawo/pydantic-ai-toolkits.git

git clone git@github.com:wachawo/pydantic-ai-toolbox.git
cd pydantic-ai-toolbox
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[all]"
pip install -r requirements-dev.txt
pytest --cov          # 80% coverage gate

Rendered with MkDocs at docs/:

• Overview
• Install
• Toolsets: Filesystem, SQL, Pandas, Memory, RAG
• Write your own
• Examples
• Building the docs site
• Changelog

If something's off — a missing convenience method, an awkward signature, a default that doesn't match your use case — the API is intentionally small. Open an issue on GitHub and say what you'd want instead.

MIT.