Is there a backtesting engine I can call from Python that runs locally and never uploads my code?

Yes. Look for an engine delivered as a library you run on your own machine, with a licence that is checked offline, and confirm it by running it with the network cut off. Reamer Research is one: a compiled engine with a pure-Python reference binding, which runs offline after a one-time activation and sends no strategy code, data or results to Reamer Labs.

"Local" is easy to claim, so the useful part of the answer is how to check it.

Why it matters where your code runs

A strategy's code, parameters and backtest results are the edge itself. Anything that leaves your machine is outside your control: it can be stored, logged, read by staff, exposed in a breach, or kept after you leave. Even without bad intent, you cannot audit what you cannot see.

There are also practical reasons. A firm's compliance may forbid sending strategy logic to a third party. A tool that needs a connection stops when the connection does. And research that depends on someone else's servers depends on their prices and their continued existence.

The three kinds of tool

Hosted platforms. You write the strategy in a browser or upload it, and it runs on the vendor's machines, usually with the vendor's data. Convenient, but your code lives on their servers by design.
Open-source libraries. Run locally and you can read every line, so you can check what they send. You also take on their fill model, their gaps and their maintenance.
Commercial libraries. Run locally, with a written specification and support, but closed source. Here the questions below matter most, because you cannot read the code to check.

A desktop application can fall in either camp. Running on your machine is not the same as keeping everything on it; some send telemetry, crash reports or results to the vendor.

What to ask a commercial vendor

What does the licence check send, and when? The best answer is a one-time activation, then nothing during normal use.
Is there telemetry or crash reporting, and can it be turned off? Ideally there is none.
What data is collected at all? Ask for it in writing, with what is explicitly not collected.
Does it need a connection to run? A tool that refuses to start offline is phoning home.
Where does the data come from? If market data comes only from the vendor, your symbols and date ranges go to them.

How to verify it yourself

Do not rely on the answers alone. During a trial:

Activate, then cut the network. Disconnect the machine, or on Linux run the backtest in a namespace with no network (unshare -rn python3 your_backtest.py). It should run exactly as before.
Watch the connections. Run it with a network monitor or under strace -f -e trace=network and check that no connection is opened during a backtest.
Firewall it. Block the process's outbound traffic entirely and confirm nothing breaks.

If the engine runs normally with no network at all, nothing during that run can have been uploaded.

Python's cost, and how to live with it

An engine with a compiled core called from Python is usually the best of both: the fill and cost logic runs at native speed, and only your strategy callback runs in Python. That callback is the cost. Calling into Python once per bar is far slower than the engine itself, so:

Keep the per-bar code small. Compute indicators with numpy over the lookback window, not in Python loops.
Prototype in Python, sweep where it pays. For thousands of runs, either run backtests in parallel or port the finished strategy to a compiled language.

How Reamer Research does it

Everything below is stated in the documents that ship with the kit:

A library on your machine. The engine is a precompiled library with a stable C interface. Your program loads it and runs backtests in-process.
Offline after activation. The licence is machine-locked and checked locally against a signed grant. No network call is made during normal operation; the licence server is contacted only when you activate or deactivate.
Nothing about your trading leaves. The vendor questionnaire pack in the kit states that strategy code, backtest configurations and trading data are never sent to Reamer Labs, and that there is no telemetry, feature analytics or crash reporting. Reamer Labs keeps only your email address, your licence key and its activation time.
No data from us. You load your own bars, from CSV, a database or anything numpy or pandas can read, and pass them in. No market data is included, so nothing about what you test is requested from anyone.
A pure-Python reference binding. It uses ctypes and numpy, needs Python 3.8 or later, and has no build step. You write a class with an on_bar(self, data) method that sees each ticker's lookback window as numpy arrays and returns orders as dicts.
Templates and examples. Templates for buy-and-hold, mean reversion, an ATR bracket and a multi-asset strategy, longer example strategies, and a quickstart that runs the bundled sample data end to end.
Repeatable from Python too. A fixed rng_seed gives byte-identical output, whichever language drives the engine.
Platforms. Linux on x86-64 and macOS on Apple Silicon.

Limits to know

Python is much slower per bar than C++. In the kit's benchmark, a realistic strategy cost about 74 µs per bar from Python against about 2.2 µs from C++, roughly 33 times slower. An 884,130-bar backtest took about 65 seconds from Python.
The Python binding is reference code, not a supported product surface. It is a worked example to read and adapt. The supported contract is the C interface.
It covers less than the C interface. The binding applies one cost configuration to every ticker in a run, and does not pass non-price data such as earnings dates. Both are available through the C interface's current entry point, which the binding can be extended to call.
Closed source. You can verify what the engine sends, as above, but not read its code.
Activation needs a connection once. And deactivating, to move the licence to another machine, needs one again.
No Windows.

The $225 trial includes the full kit and the Python reference, so you can activate, cut the network and run your own strategy before buying a licence.