# Marcus Delgado, Senior Backend Engineer at Vantix Payments — read of Codex Local, June 16 2026

> 9 years writing Python and Go at fintech startups, currently the loudest voice in Slack whenever InfoSec sends one of those "we need to talk about Copilot" emails.

## How I got here

Our InfoSec lead forwarded a memo two weeks ago asking us to document what code Copilot is sending to Microsoft. I don't think it's a big deal personally, but I started searching "copilot alternative local model vscode" out of curiosity and this page showed up third or fourth result. No ad label, no LinkedIn, just a search. Clicked it on my lunch break.

## What I clicked first

The hero actually got me. "Run AI Code Completions Offline. Zero Subscriptions. Zero Telemetry." That's three things that directly answer the three questions on my InfoSec lead's memo. Not fluffy. Not "the future of coding is here." Just three claims that map to three real objections I have open in a browser tab.

The system requirements table also made me stop in a good way. "NVIDIA RTX 3060+ (6GB VRAM)... Completions in 50-100ms" versus "16GB RAM, modern CPU... Completions in 500ms-2s." Someone thought about this. They didn't just say "works on your machine."

## Where I paused

The install steps. They list the default endpoint as `localhost:5000` in one place and then the actual Ollama code block uses `localhost:11434`. That's a small thing but it's the kind of thing that happens when documentation is written by someone who set it up once and then wrote the steps from memory. I've done it myself. It's not a dealbreaker but I noticed it.

I also paused on "sub-100ms completions" in the features list. That's specific enough to be testable. I appreciate that they didn't say "blazing fast." But I'd want to know: 100ms on what hardware, for what context length, with what model? Because CodeLlama 13B on an RTX 3060 pulling 1000 tokens of context is not the same problem as 100 tokens on an A100.

## What I distrusted

The bottom of the page. After reading what felt like a real product, I hit: "Adopt this idea. Browse free. Unlock for $5. Adopt for $99." And then "Estimates only - no live customer revenue claimed."

So this isn't a product. It's a product concept being sold as a kit. "Built with love at Wishdeal Studio" is a studio that packages startup ideas. The GitHub link in the footer doesn't go to a repo with actual stars or actual commits. There's no download count. No "X installs from the marketplace." The VSCode marketplace link in the hero is presumably dead or points to a placeholder.

I spent four minutes reading what I thought was a real extension and it turns out I was reading a pitch deck formatted as a landing page. That's the kind of thing that makes me feel like I've been lightly deceived, even if the intent wasn't malicious.

"Works Like Copilot" also landed flat. That's the one phrase where the copy stopped thinking and just reached for the comparison. Everything else on the page avoids that trap and then it slips in anyway.

## What would convince me

If this were a real product: a link to the actual VSCode marketplace listing with an install count above 1,000. A GitHub repo with at least 6 months of commit history and open issues that look like real user bugs, not placeholder TODOs. One screen recording of someone using it on a real codebase, not a demo with `hello_world.py`.

Specifically on the privacy claim: a one-paragraph technical writeup of how completions are routed. Not a bullet point saying "code never leaves your machine." Show me the network calls. Show me the Wireshark capture. That would actually move me.

## What I'd ask in an email reply

1. Is there a working VSCode extension I can install right now, or is this a build kit for someone who wants to build this product?

2. The page mentions "No telemetry" but the extension itself would presumably run inside VSCode. Does it register any telemetry with the VSCode extension host, and if not, how do I verify that from the source?

3. The recommended model is CodeLlama 13B. Have you benchmarked it against the current Ollama default models on HumanEval or any actual completion accuracy benchmark, or is that recommendation based on community consensus from 2023?

## Verdict: dismissive

Not because the idea is bad. The idea is legitimately good and the copy is cleaner than 80% of dev tools I've seen. But the "Estimates only - no live customer revenue claimed" disclosure at the bottom means I'm reading a concept, not a product. I don't have time to adopt someone else's build kit for a tool that Cursor and Continue.dev already ship in working form.

---
*Memo by skeptic persona, generated 2026-06-16. Studio breaks own self-grading loop.*