AI Agent 安全 Python 执行器 7.5k star 杀手级开源神器

https://github.com/pydantic/monty

这就是 Monty，Pydantic 官方出品的 Rust 极致轻量安全 Python 解释器，专为 AI Agent 设计！一句话干翻传统沙箱痛点：用 Rust 实现微秒级启动（<1μs）、近原生 CPython 性能，却比 Docker 轻量上千倍，同时严格限制 AI 生成代码对主机文件系统、网络、环境变量的访问，只允许白名单 host 函数调用，真正实现又快又安全地让 Agent 执行自己写的 Python 代码！

AI Agent 开发者、需要安全执行 LLM 生成代码的团队、Pydantic AI / Claude Code / Cursor 重度用户、追求极致性能和安全性的极客的终极执行神器！ 再也不用为了跑一段 Agent 写的代码而拉一个完整 Docker 容器、等半天启动、还担心安全问题了 用了就回不去了

pydantic/monty

Source: https://github.com/pydantic/monty

Experimental - This project is still in development, and not ready for the prime time.

A minimal, secure Python interpreter written in Rust for use by AI.

Monty avoids the cost, latency, complexity and general faff of using a full container based sandbox for running LLM generated code.

Instead, it lets you safely run Python code written by an LLM embedded in your agent, with startup times measured in single digit microseconds not hundreds of milliseconds.

What Monty can do:

• Run a reasonable subset of Python code - enough for your agent to express what it wants to do
• Completely block access to the host environment: filesystem, env variables and network access are all implemented via external function calls the developer can control
• Call functions on the host - only functions you give it access to
• Run typechecking - monty supports full modern python type hints and comes with ty included in a single binary to run typechecking
• Be snapshotted to bytes at external function calls, meaning you can store the interpreter state in a file or database, and resume later
• Startup extremely fast (<1μs to go from code to execution result), and has runtime performance that is similar to CPython (generally between 5x faster and 5x slower)
• Be called from Rust, Python, or Javascript - because Monty has no dependencies on cpython, you can use it anywhere you can run Rust
• Control resource usage - Monty can track memory usage, allocations, stack depth, and execution time and cancel execution if it exceeds preset limits
• Collect stdout and stderr and return it to the caller
• Run async or sync code on the host via async or sync code on the host
• Use a small subset of the standard library: sys, os, typing, asyncio, re, datetime, json, dataclasses (soon)

What Monty cannot do:

• Use the rest of the standard library
• Use third party libraries (like Pydantic), support for external python library is not a goal
• define classes (support should come soon)
• use match statements (again, support should come soon)

In short, Monty is extremely limited and designed for one use case:

To run code written by agents.

For motivation on why you might want to do this, see:

• Codemode from Cloudflare
• Programmatic Tool Calling from Anthropic
• Code Execution with MCP from Anthropic
• Smol Agents from Hugging Face

In very simple terms, the idea of all the above is that LLMs can work faster, cheaper and more reliably if they’re asked to write Python (or Javascript) code, instead of relying on traditional tool calling. Monty makes that possible without the complexity of a sandbox or risk of running code directly on the host.

Note: Monty will (soon) be used to implement codemode in Pydantic AI

Usage

Monty can be called from Python, JavaScript/TypeScript or Rust.

Python

To install:

uv add pydantic-monty

(Or pip install pydantic-monty for the boomers)

Usage:

from typing import Any

import pydantic_monty

code = """
async def agent(prompt: str, messages: Messages):
    while True:
        print(f'messages so far: {messages}')
        output = await call_llm(prompt, messages)
        if isinstance(output, str):
            return output
        messages.extend(output)

await agent(prompt, [])
"""

type_definitions = """
from typing import Any

Messages = list[dict[str, Any]]

async def call_llm(prompt: str, messages: Messages) -> str | Messages:
    raise NotImplementedError()

prompt: str = ''
"""

m = pydantic_monty.Monty(
    code,
    inputs=['prompt'],
    script_name='agent.py',
    type_check=True,
    type_check_stubs=type_definitions,
)


Messages = list[dict[str, Any]]


async def call_llm(prompt: str, messages: Messages) -> str | Messages:
    if len(messages) < 2:
        return [{'role': 'system', 'content': 'example response'}]
    else:
        return f'example output, message count {len(messages)}'


async def main():
    output = await m.run_async(
        inputs={'prompt': 'testing'},
        external_functions={'call_llm': call_llm},
    )
    print(output)
    #> example output, message count 2


if __name__ == '__main__':
    import asyncio

    asyncio.run(main())

Iterative Execution with External Functions

Use start() and resume() to handle external function calls iteratively, giving you control over each call:

import pydantic_monty

code = """
data = fetch(url)
len(data)
"""

m = pydantic_monty.Monty(code, inputs=['url'])

# Start execution - pauses when fetch() is called
result = m.start(inputs={'url': 'https://example.com'})

print(type(result))
#> <class 'pydantic_monty.FunctionSnapshot'>
print(result.function_name)  # fetch
#> fetch
print(result.args)
#> ('https://example.com',)

# Perform the actual fetch, then resume with the result
result = result.resume({'return_value': 'hello world'})

print(type(result))
#> <class 'pydantic_monty.MontyComplete'>
print(result.output)
#> 11

Serialization

Both Monty and snapshot types like FunctionSnapshot can be serialized to bytes and restored later. This allows caching parsed code or suspending execution across process boundaries:

import pydantic_monty

# Serialize parsed code to avoid re-parsing
m = pydantic_monty.Monty('x + 1', inputs=['x'])
data = m.dump()

# Later, restore and run
m2 = pydantic_monty.Monty.load(data)
print(m2.run(inputs={'x': 41}))
#> 42

# Serialize execution state mid-flight
m = pydantic_monty.Monty('fetch(url)', inputs=['url'])
progress = m.start(inputs={'url': 'https://example.com'})
state = progress.dump()

# Later, restore and resume (e.g., in a different process)
progress2 = pydantic_monty.load_snapshot(state)
result = progress2.resume({'return_value': 'response data'})
print(result.output)
#> response data

Rust

use monty::{MontyRun, MontyObject, NoLimitTracker, PrintWriter};

let code = r#"
def fib(n):
    if n <= 1:
        return n
    return fib(n - 1) + fib(n - 2)

fib(x)
"#;

let runner = MontyRun::new(code.to_owned(), "fib.py", vec!["x".to_owned()]).unwrap();
let result = runner.run(vec![MontyObject::Int(10)], NoLimitTracker, PrintWriter::Stdout).unwrap();
assert_eq!(result, MontyObject::Int(55));

Serialization

MontyRun and RunProgress can be serialized using the dump() and load() methods:

use monty::{MontyRun, MontyObject, NoLimitTracker, PrintWriter};

// Serialize parsed code
let runner = MontyRun::new("x + 1".to_owned(), "main.py", vec!["x".to_owned()]).unwrap();
let bytes = runner.dump().unwrap();

// Later, restore and run
let runner2 = MontyRun::load(&bytes).unwrap();
let result = runner2.run(vec![MontyObject::Int(41)], NoLimitTracker, PrintWriter::Stdout).unwrap();
assert_eq!(result, MontyObject::Int(42));

PydanticAI Integration

Monty will power code-mode in Pydantic AI. Instead of making sequential tool calls, the LLM writes Python code that calls your tools as functions and Monty executes it safely.

import asyncio
import json

import logfire
from httpx import AsyncClient
from pydantic_ai import Agent, RunContext
from pydantic_ai.toolsets.code_mode import CodeModeToolset
from pydantic_ai.toolsets.function import FunctionToolset
from typing_extensions import TypedDict

logfire.configure()
logfire.instrument_pydantic_ai()


class LatLng(TypedDict):
    lat: float
    lng: float


weather_toolset: FunctionToolset[AsyncClient] = FunctionToolset()


@weather_toolset.tool
async def get_lat_lng(
    ctx: RunContext[AsyncClient], location_description: str
) -> LatLng:
    """Get the latitude and longitude of a location."""
    # NOTE: the response here will be random, and is not related to the location description.
    r = await ctx.deps.get(
        'https://demo-endpoints.pydantic.workers.dev/latlng',
        params={'location': location_description},
    )
    r.raise_for_status()
    return json.loads(r.content)


@weather_toolset.tool
async def get_temp(ctx: RunContext[AsyncClient], lat: float, lng: float) -> float:
    """Get the temp at a location."""
    # NOTE: the responses here will be random, and are not related to the lat and lng.
    r = await ctx.deps.get(
        'https://demo-endpoints.pydantic.workers.dev/number',
        params={'min': 10, 'max': 30},
    )
    r.raise_for_status()
    return float(r.text)


@weather_toolset.tool
async def get_weather_description(
    ctx: RunContext[AsyncClient], lat: float, lng: float
) -> str:
    """Get the weather description at a location."""
    # NOTE: the responses here will be random, and are not related to the lat and lng.
    r = await ctx.deps.get(
        'https://demo-endpoints.pydantic.workers.dev/weather',
        params={'lat': lat, 'lng': lng},
    )
    r.raise_for_status()
    return r.text


agent = Agent(
    'gateway/anthropic:claude-sonnet-4-5',
    # toolsets=[weather_toolset],
    toolsets=[CodeModeToolset(weather_toolset)],
    deps_type=AsyncClient,
)


async def main():
    async with AsyncClient() as client:
        await agent.run('Compare the weather of London, Paris, and Tokyo.', deps=client)


if __name__ == '__main__':
    asyncio.run(main())

Community Bindings

• Go: gomonty - Go bindings for the Monty interpreter
• Dart/Flutter: dart_monty (github) (pub.dev)- Dart/Flutter bindings for Monty

Alternatives

There are generally two responses when you show people Monty:

1. Oh my god, this solves so many problems, I want it.
2. Why not X?

Where X is some alternative technology. Oddly often these responses are combined, suggesting people have not yet found an alternative that works for them, but are incredulous that there’s really no good alternative to creating an entire Python implementation from scratch.

I’ll try to run through the most obvious alternatives, and why there aren’t right for what we wanted.

NOTE: all these technologies are impressive and have widespread uses, this commentary on their limitations for our use case should not be seen as a criticism. Most of these solutions were not conceived with the goal of providing an LLM sandbox, which is why they’re not necessary great at it.

See ./scripts/startup_performance.py for the script used to calculate the startup performance numbers.

Details on each row below:

Monty

• Language completeness: No classes (yet), limited stdlib, no third-party libraries
• Security: Explicitly controlled filesystem, network, and env access, strict limits on execution time and memory usage
• Start latency: Starts in microseconds
• Setup complexity: just pip install pydantic-monty or npm install @pydantic/monty, ~4.5MB download
• File mounting: Strictly controlled, see #85
• Snapshotting: Monty’s pause and resume functionality with dump() and load() makes it trivial to pause, resume and fork execution

Docker

• Language completeness: Full CPython with any library
• Security: Process and filesystem isolation, network policies, but container escapes exist, memory limitation is possible
• Start latency: Container startup overhead (~195ms measured)
• Setup complexity: Requires Docker daemon, container images, orchestration, python:3.14-alpine is 50MB - docker can’t be installed from PyPI
• File mounting: Volume mounts work well
• Snapshotting: Possible with durable execution solutions like Temporal, or snapshotting an image and saving it as a Docker image.

Pyodide

• Language completeness: Full CPython compiled to WASM, almost all libraries available
• Security: Relies on browser/WASM sandbox - not designed for server-side isolation, python code can run arbitrary code in the JS runtime, only deno allows isolation, memory limits are hard/impossible to enforce with deno
• Start latency: WASM runtime loading is slow (~2800ms cold start)
• Setup complexity: Need to load WASM runtime, handle async initialization, pyodide NPM package is ~12MB, deno is ~50MB - Pyodide can’t be called with just PyPI packages
• File mounting: Virtual filesystem via browser APIs
• Snapshotting: Possible with durable execution solutions like Temporal presumably, but hard

starlark-rust

See starlark-rust.

• Language completeness: Configuration language, not Python - no classes, exceptions, async
• Security: Deterministic and hermetic by design
• Start latency: runs embedded in the process like Monty, hence impressive startup time
• Setup complexity: Usable in python via starlark-pyo3
• File mounting: No file handling by design AFAIK?
• Snapshotting: Impossible AFAIK?

WASI / Wasmer

Running Python in WebAssembly via Wasmer.

• Language completeness: Full CPython, pure Python external packages work via mounting, external packages with C bindings don’t work
• Security: In principle WebAssembly should provide strong sandboxing guarantees.
• Start latency: The wasmer python package hasn’t been updated for 3 years and I couldn’t find docs on calling Python in wasmer from Python, so I called it via subprocess. Start latency was 66ms.
• Setup complexity: wasmer download is 100mb, the “python/python” package is 50mb.
• FOSS: I marked this as “free *” since the cost is zero but not everything seems to be open source. As of 2026-02-10 the python/python wasmer package package has no readme, no license, no source link and no indication of how it’s built, the recently uploaded versions show size as “0B” although the download is ~50MB - the build process for the Python binary is not clear and transparent. (If I’m wrong here, please create an issue to correct correct me)
• File mounting: Supported
• Snapshotting: Supported via journaling

sandboxing service

Services like Daytona, E2B, Modal.

There are similar challenges, more setup complexity but lower network latency for setting up your own sandbox setup with k8s.

• Language completeness: Full CPython with any library
• Security: Professionally managed container isolation
• Start latency: Network round-trip and container startup time. I got ~1s cold start time with Daytona EU from London, Daytona advertise sub 90ms latency, presumably that’s for an existing container, not clear if it includes network latency
• FOSS: Pay per execution or compute time, some implementations are open source
• Setup complexity: API integration, auth tokens - fine for startups but generally a non-start for enterprises
• File mounting: Upload/download via API calls
• Snapshotting: Possible with durable execution solutions like Temporal, also the services offer some solutions for this, I think based con docker containers

YOLO Python

Running Python directly via exec() (~0.1ms) or subprocess (~30ms).

• Language completeness: Full CPython with any library
• Security: None - full filesystem, network, env vars, system commands
• Start latency: Near-zero for exec(), ~30ms for subprocess
• Setup complexity: None
• File mounting: Direct filesystem access (that’s the problem)
• Snapshotting: Possible with durable execution solutions like Temporal

GitTrend (@GitTrend0x): AI Agent 专用轻量级无头浏览器 14.9k star 杀手级开源神器

https://t.co/Dk8YYYVGda

这就是 Obscura，Rust 编写的下一代无头浏览器，专为 AI Agent 和大规模爬虫打造！一句话干翻传统 headless Chrome：5-10 倍更快、内存占用极低（仅 30MB）、原生 V8 引擎 + 内置顶级反指纹，支持完整 CDP