agex (a portmanteau of agent execution) is a Python library for building AI agents that work directly with your code.
You define a typed Python function with @task and the agent fills it in. It writes sandboxed Python that calls into the modules you've whitelisted, returning real Python objects (DataFrames, Plotly figures, your Pydantic models) that flow back into your code unchanged. There's no JSON serialization at the boundary and no separate runtime to deploy: agex runs inside your application's Python process.
Because the sandbox is pure-Python AST rewriting, the same agent code runs in-process, in a subprocess, in a kernel-isolated worker, or (via Pyodide) entirely in a browser tab. agex-studio is the proof-of-concept: pandas, scikit-learn, plotly, and a chat agent all running client-side with no backend.
import pandas as pd
from agex import Agent
agent = Agent()
agent.module(pd)
@agent.task
def summarize(df: pd.DataFrame) -> dict[str, float]:
"""Return summary statistics for the numeric columns."""
pass
stats = summarize(my_dataframe) # real dict[str, float]
- Typed function tasks -
@taskdeclares the input/output contract; the agent fulfills it. - Curated Python environment - whitelist exactly which modules and classes the agent can use, with per-member visibility.
- Versioned workspace - virtual filesystem, the agent's session memory, and event log are all kvgit-backed, with checkpoints and time-travel.
- Multi-agent orchestration - coordinate agents with regular Python control flow; sub-agents are just functions.
- Flexible execution - in-process by default; subprocess, kernel-isolated, browser (Pyodide), or remote (Modal, HTTP) when you need them.
For a deeper dive, see the agex101 tutorial or the geospatial routing example for multi-library integration. For a NiceGUI integration demo, see agex-ui.
Complete documentation is hosted at ashenfad.github.io/agex.
Key sections:
Install agex with your preferred LLM provider:
pip install "agex[openai]" # OpenAI models
pip install "agex[anthropic]" # Anthropic Claude models
pip install "agex[gemini]" # Google Gemini models
# Or with all providers
pip install "agex[all-providers]"This is a hobby project in active development. The core concepts are stabilizing but the API should be considered experimental.
If you're looking for a more battle-tested library built on the same "agents-that-think-in-code" idea, smolagents (Hugging Face) is the closest cousin and a good choice. agex explores a different shape: an embeddable library you import into your application, with typed function contracts and a pure-Python sandbox that runs anywhere Python runs (including the browser).
agex is composed of several focused libraries that can also be used independently:
| Library | Purpose |
|---|---|
| sandtrap | In-process Python sandbox via AST rewriting |
| kvgit | Versioned key-value store with git-like semantics |
| monkeyfs | Filesystem interception via monkey-patching |
| termish | Virtual terminal with shell-like commands |
| reprobate | Budget-controlled repr for Python objects |
Bug reports, ideas, and pull requests welcome - see GitHub Issues or the Contributing Guide.