Agents increasingly answer by writing code: a data transform, a chart, a loop that calls your tools in parallel instead of one JSON call at a time. That code is untrusted by construction. pydeno runs it behind a boundary that holds even if the code is hostile, and gives it exactly the tools you pass in.

from pydeno import Pydeno

with Pydeno() as pool, pool.checkout() as session:        # OS sandbox required, limits on, warm worker
    session.feed_run("const prices = {A1: 3.5}")
    session.feed_run("prices[sku] * await rate()",         # 7.0: state persists between feeds
                     inputs={"sku": "A1"}, external_lookup={"rate": lambda: 2})
    session.feed_run("new Array(2 ** 32 - 1).fill(0)")     # PydenoCrashedError. Your process is fine.

If you know Monty, you already know Pydeno: the same Pydeno / checkout / feed_run / feed_start / dump shape, for JavaScript. Where Monty runs a Python subset, pydeno runs real, modern JavaScript (V8, the engine behind Chrome and Node), so model-written code behaves like JavaScript and real libraries (Vega, ECharts, three.js, d3, pptxgenjs) run unmodified.

Which one? Use Monty by default: for plain logic it is smaller, starts in microseconds and needs no process. Reach for pydeno when the code needs the JavaScript ecosystem: rich UI, charts, graphs, 3D and documents. They also work side by side (see the comparison).

Quickstart

pip install pydeno     # or: uv pip install pydeno     (Python 3.10+, macOS or Linux)

from pydeno import Pydeno

with Pydeno() as pool:
    with pool.checkout() as session:
        session.feed_run("const x = 20")
        print(session.feed_run("x + 1"))                                                          # 21
        print(session.feed_run("await lookup(7)", external_lookup={"lookup": lambda i: i * 6}))   # 42

from pydeno import AsyncPydeno

async with AsyncPydeno() as pool:
    async with pool.checkout() as session:
        await session.feed_run("1 + 1")                                                           # 2

Defaults: sandbox="require" (pydeno.sandbox_status() explains a refusal), jitless V8, host errors redacted, 30 s per feed, 512 MiB, 1000 external calls per session, and a worker never serves two sessions. The front-door guide maps every Monty name and limit.

One door, the same names as Monty. Pydeno and AsyncPydeno are the default path for everything; you should not need another class to run untrusted code.

The front-door guide maps every Monty name and limit. The lower-level classes (IsolatedRuntime, AgentSandbox, SandboxPool) stay available under Which runtime?.

From a shell, the same sandboxed worker (command line guide):

pydeno '[1, 2, 3].map(x => x * 2)'     # prints [2, 4, 6]; exit code 1 on a JavaScript error

Code mode for AI agents

Code mode lets the model write one program that calls your tools, instead of one tool call per turn. pydeno gives that program a safe place to run.

from pydeno import Pydeno

TOOLS = {"search_orders": search_orders, "get_customer": get_customer}      # your Python functions

with Pydeno() as pool, pool.checkout(limits={"max_feed_duration_secs": 10, "max_memory": 256 * 2**20}) as s:
    answer = s.feed_run(
        model_written_javascript,       # e.g. loops over orders, calls get_customer in Promise.all, aggregates
        external_lookup=TOOLS,          # the ONLY way out, with a total call budget
        print_callback=lambda stream, text: log(stream, text),
    )

• Approval flows: feed_start returns a snapshot at every tool call; inspect snapshot.function_name and snapshot.args, then resume(value=...) or resume(error=...). dump() it and continue in another process.
• Typed declarations for your prompt: your tools become tools.* functions with TypeScript stubs; JSON-Schema tools and a lazy tool catalog keep the prompt small even with hundreds of tools.
• Agent frameworks: JSCodeMode is the JavaScript counterpart of pydantic-ai's Monty-based code mode, and an allow-listed, SSRF-safe http_fetch tool is included.
• Results you can bound: execute() returns an ExecutionResult (stdout, stderr, result, error) with ordered, size-capped console output, so one runaway console.log cannot flood your logs or your context window.

Security

pydeno assumes the code is hostile and the engine can have bugs. Each layer assumes the one above has failed. The threat model is in SECURITY.md; the details are in the isolation guide.

How we check it.

• A battery of hostile-guest probes (scripts/autoresearch/metric_security.py): poisoned prototypes, forged frames, deadline bypasses, memory bombs, state leaking between commands, console escape injection. The goal is zero, and every newly found attack class becomes a probe first.
• Independent review of every change that touches the worker, wire, bridge, supervision, sandbox policy or journals. Reviews have repeatedly found real defects, including in the fixes themselves, which is why they are mandatory. The findings table, misses included, is the security report.
• Assume-breach syscall sweeps against the real kernel, a verified seccomp program, fuzzing, hostile-worker fakes, and Linux on many distros and both CPU architectures (details below).

What we do not claim. A V8 bug is contained in the worker, not prevented. V8 is a large engine we cannot audit end to end, which is why the layers around it exist. We also cannot lock the engine down completely: real libraries such as three.js, Vega and ECharts need much of the language, so a restriction becomes a default only when those libraries still run (strict_eval stays opt-in for that reason). The containment is the process and the OS sandbox, not a minimal engine; there has been no independent outside security review yet (it is planned before 1.0, see the roadmap). The in-process Runtime is not safe for hostile code: use Pydeno / IsolatedRuntime. Windows has no isolated worker. For multi-tenant use, put the whole process in a locked-down container or microVM (deployment guidance). Found a way out? Please report it privately, as described in SECURITY.md.

Why pydeno is harder to break than most sandboxes

Running untrusted JavaScript has a long history of sandboxes that were broken. The same few design mistakes keep appearing, and each one is something pydeno was built to avoid:

What this does not mean. pydeno is not unbreakable. It contains the engine, it does not shrink it: a V8 bug is contained in the worker, not prevented, and V8 is large. A small interpreter with a minimal trusted base has a real advantage for code that is mostly Python logic. pydeno has had no outside security review yet; it has had an independent review of every change that touches the boundary, and that process has repeatedly found real defects, including in fixes. We publish the misses as well as the catches in the security report.

Speed

Measured on macOS arm64 with release builds, using paired A/B runs; reproduce with benches_py/alternatives_bench.py and scripts/autoresearch/metric_speed.py. Numbers vary with the machine.

Jitless V8 (the default) makes compute-heavy code roughly 1.5 to 8 times slower than with the JIT; jitless=False trades that back for a larger attack surface, and still runs behind the full OS sandbox.

How it compares

Monty is the better choice when the model's code is mostly Python logic and you want the smallest trusted base; pydeno is the choice when you need JavaScript and its ecosystem. They combine well (see the showcase). The full, frank comparison, including what pydeno cannot do and measured numbers, is on the alternatives page.

Advanced: the building blocks

Pydeno is built from these, which stay available unchanged.

Tools the guest can call, with a total budget:

from pydeno import IsolatedRuntime, ToolBridge

bridge = ToolBridge({"get_weather": get_weather, "send_email": send_email}, max_calls=50)

with IsolatedRuntime(sandbox="require") as rt:
    bridge.attach(rt)
    rt.eval("tools.get_weather('Zurich')")

The call that exceeds max_calls never reaches your Python function; the guest gets a catchable ToolBudgetError. A Python exception inside a tool reaches JavaScript as a real Error whose name is the exception's class, so the model's code can branch on it. Host errors are redacted by default.

Make the easy path the safe one (the module-level helpers then use a sandboxed worker):

import pydeno

pydeno.configure_default_runtime(isolated=True, sandbox="require")
pydeno.eval("1 + 1")

Run real libraries. Many assume browser basics a bare isolate lacks. Opt in to a small, pure-JavaScript set (virtual-time timers, TextEncoder, btoa, Blob, EventTarget, AbortController, structuredClone) that never defines window or document:

from pydeno import IsolatedRuntime, RuntimeConfig, WEB_POLYFILLS

rt = IsolatedRuntime(RuntimeConfig(bootstrap=WEB_POLYFILLS))

Async code: AsyncPydeno, AsyncIsolatedRuntime and AsyncAgentSandbox never block the event loop and use no thread per runtime.

Showcase: Python and JavaScript, both sandboxed

A model that writes code wants the best language for each half of the job: Python for wrangling data, JavaScript for what only the JS ecosystem does well. Pair pydeno with Monty (Pydantic's Python sandbox) and run both without trusting either, sharing one set of host tools and one call budget:

with Monty() as pool, pool.checkout() as py:                      # the model's Python, in Monty
    analysis = py.feed_run(model_python, external_lookup={"fetch_sales": fetch_sales})

with IsolatedRuntime(RuntimeConfig(bootstrap=WEB_POLYFILLS), sandbox="require") as js:
    js.bind_function("getAnalysis", lambda: analysis)             # hand the result across
    js.bind_function("fetch_sales", fetch_sales)                  # the same tool, same budget
    svg = await js.eval_async(model_javascript)                   # the model's JS, in pydeno

1. Python half, in Monty
  [python] best region: East (1750)
  [python] refused: PermissionError: Permission denied: '/etc/passwd'
2. JavaScript half, in pydeno
  [js]     refused: Evaluation failed: ReferenceError: fetch is not defined
  [js]     sandbox: seatbelt
3. wrote sales.svg (11021 bytes); tool calls used: 2 of 5

Python crunched the numbers; JavaScript drew the chart with Vega-Lite; each sandbox refused its own attempt to reach outside, and both drew on the same five-call tool budget. Full runnable example: examples/monty_and_pydeno.py.

More of the same: Python prepares, JavaScript builds

Monty is excellent at the data half, and the JavaScript ecosystem has libraries nothing in Python matches. Each example below is a complete, tested program: the model's Python runs in Monty, the model's JavaScript runs in pydeno, neither can reach your files, network or environment, and each test checks the answer against an independent computation (and Monty's against CPython's).

How fast are they together? One warm worker, median of three, on an Apple-silicon Mac, with V8 in its secure default (jitless) and with the JIT turned on:

Most turns finish in well under a second even in the secure mode; Monty's data half stays within about 1.1x of plain CPython. The JIT matters for compute-heavy JavaScript (layouts, raycasting), and it is exactly the part of V8 where most exploits live, which is why it is off by default. If you trust the code a little more, IsolatedRuntime(jitless=False) buys the second column and still runs behind the full OS sandbox. Reproduce with benches_py/monty_three_bench.py.

How it works

flowchart LR
    app["Your Python app"] -- "eval, tools" --> parent["IsolatedRuntime<br/>(supervisor)"]
    parent <-->|"validated frames only"| worker
    subgraph worker["Worker process: sandboxed before V8 starts"]
        v8["V8 (jitless)<br/>guest JavaScript"]
    end
    v8 -. "bound functions only" .-> parent
    parent --> tools["Your tools"]

The worker is a separate process with an empty environment, its own session, no privileges, and an OS sandbox applied before the JavaScript engine exists. The parent enforces the deadline, memory and call budgets from outside and treats everything the worker sends as untrusted input.

Which runtime?

How it was tested

The sandbox is tested the way an attacker would try it: from inside, and against the real kernel.

• Assume-breach syscall sweeps. A sandboxed process fires every syscall in the kernel's table with garbage arguments, from a fresh process each time, and records what is still reachable. Dangerous calls are fired for real (inside a container, never on a host).
• The seccomp program is verified, not trusted. A BPF interpreter in the tests runs the filter's decision logic on any platform, and every syscall number in it is checked against the Linux kernel's own tables (regenerated from a pinned kernel tag by scripts/gen_syscall_tables.py).
• A capability checklist. What untrusted code tries first (read or write a file, reach the network, spawn a process, read the environment; Deno, Node and browser globals; the runtime's own plumbing), through both eval and eval_async, plus a check that nothing reached the host.
• Hostile peers. Fake workers that stop reading, send malformed or oversized frames, flood hash tables or lie about types; fuzzing with Hypothesis; and a port of pydantic/monty's security suite (the cases an in-process runtime cannot survive are pinned as strict expected failures).
• Two codecs, tested against each other. The wire codec is native (Rust) with a readable Python reference; hundreds of tests, including generated frames, require identical results and errors.
• Real libraries, inside the sandbox. pptxgenjs, three.js (GLB export), Vega-Lite and dagre give the same results as an unsandboxed runtime, with their bytes pinned in vendor/libs/; about 25 more (d3, ECharts, jsPDF, Tailwind v4, Prettier, ...) were checked by hand.
• Many kernels. The Linux suite runs on 12 distro images (Debian, Ubuntu, Fedora, AlmaLinux, Amazon Linux; Python 3.10 to 3.14) on both x86_64 and aarch64, and under simulated kernels that lack Landlock, seccomp or both, to prove the sandbox degrades honestly and sandbox="require" fails closed. See .github/workflows/test.yml.
• Zero skips. Platform-specific tests are deselected, never skipped; CI enforces a skip budget of zero so an unrun test cannot hide.
• The worker proves its own confinement. Before any guest code exists it tries to read a file, write one, spawn a process, connect out, signal its parent and (on macOS) read the parent's environment and the machine's hardware ID; a complete sandbox that lets one through refuses to start. Each probe is also checked in reverse: it must report a breach in an unsandboxed process.
• Independent review, and what it found. Three AI reviewers (each given a separate slice of the sandbox and told to reproduce before reporting), GitHub Copilot, and research into how other sandboxes were attacked. The result is a findings table that lists the misses as well as the catches, including a macOS leak of the host's environment, a bridge bug that let a guest abort the process, and a V8 x86_64 startup trap that only a native x86_64 run revealed. Read it in the security report.
• An autonomous pentest agent, run against the source. Before 0.8 we ran Strix (an open-source AI pentest agent) as a white-box review of a clean export of the code, in standard mode, locally, with no outside host reachable and a hard budget cap. It found no host escape and no capability-token bypass in the paths it reviewed. What it did report went through the same path as every other review finding: a failing test first, the smallest fix, a review by a different model. Its findings are not published; that is deliberate for a public repository. It is one more reviewer, not a proof: it read the code and did not exercise the native layer, so the OS-level checks above still carry that part. It is set up to run on a schedule (how).
• A probe battery that must stay at zero. scripts/autoresearch/metric_security.py runs hostile guests and hostile callers (limits, tampered binds, forged journals, oversized values, terminal escapes) and has to report zero violations before anything merges to the release branch.

Integrations

• FastMCP tool bridge: expose FastMCP tools to sandboxed JS via bind_function and an in-process fastmcp.Client
• pydantic-ai code mode (JSCodeMode): the JavaScript counterpart of pydantic-ai's Monty-based code mode. The agent gets one run_javascript tool; your other tools become typed tools.* functions the model's code calls with await and Promise.all, with retries, usage limits and approvals mapped onto pydantic-ai's own. Runs offline: examples/pydantic_ai_agent.py. pip install "pydeno[pydantic-ai]"
• llm plugin (llm-pydeno): a PyDeno toolbox for the llm CLI. One run_javascript tool runs the model's code in a sandboxed session that keeps state between calls and returns stdout, stderr, result and error with size caps. A separate package; pydeno does not depend on llm.
• Agent sessions (AgentSandbox): state across turns, pause and resume at every tool call (approval flows), a signed replay journal you can dump() and load(), and the tool descriptions and .d.ts for your prompt
• ToolBridge: several Python tools with a total call budget, typed errors the model's JS can branch on, and console.log routed back to Python
• Monty + pydeno: the model's Python runs in Monty, its JavaScript in pydeno, both sandboxed, sharing one tool and one call budget; Python computes, a Vega-Lite chart is drawn in JS
• Vendored npm libraries: run real npm document-generation libraries (pptxgenjs, pdf-lib) from their browser bundles inside the sandbox; see the guide
• Arrow IPC dataframes: move 100k+ row tables into the sandbox as Arrow IPC bytes instead of JSON objects (5 ms vs 253 ms); see the guide

Documentation

• Security policy and threat model · the isolation guide
• Quick Start · Concepts · Guides · Use cases · API reference
• Agent skill: an Agent Skills SKILL.md that teaches coding agents (Claude Code and others) to use pydeno correctly. Copy skills/pydeno/ into your agent's skills directory, e.g. .claude/skills/
• Benchmarks · Changelog

Status and contributing

pydeno is experimental: expect breaking changes between versions. Found a way out of the sandbox? Please report it privately, as described in SECURITY.md. Everything else: issues and pull requests are welcome; start with docs/contributing/.

Licensed under the MIT License.

Acknowledgements

pydeno began as a fork of jsrun by Xin Fu, which had the original idea of a Python library that runs JavaScript on a Rust runtime built on deno_core. We took that idea and its foundations, then changed a great deal: the sandboxed worker process and OS confinement, the tool boundary, the wire codec, the resource limits, and most of the tests are new. Thank you to jsrun for the starting point. The MIT licence and original copyright are kept, and docs/contributing/upstream-divergence.md records what came from where.

Thanks also to Monty by Pydantic, whose design for running agent-written code (limits, host functions, a public challenge to break it) shaped how we think about this problem; the two sandboxes work well side by side.