Developer Builds Discord Alternative With AI: Memory Leak Cut From 305.8 MB to 30.3 MB
A developer handed nearly all the coding of Nimbo, a free Discord replacement for gaming with friends, to AI, and the AI never noticed the memory leaks. After five rounds of switching the camera and screen sharing on and off, the Windows client held 305.8 MB of RAM even though the video was already off. After the fix, it held 30.3 MB. The GTCRN noise suppressor muted the crunch of an apple by 21.29 dB, while the previous RNNoise filter managed only 0.35 dB. The takeaway: AI assembles a skeleton quickly, but you have to check for leaks and audio quality by hand.
AI-processed from Habr AI; edited by Hamidun News
Imagine this: you switch your camera or screen sharing on and off five times during a call, the video stopped long ago, and the program still occupies 305.8 MB of RAM. That is exactly how the Windows client of Nimbo behaved.
Nimbo is a free voice chat for gaming with friends, and its author wrote almost all of it with the help of AI. The cause turned out to be mundane: the encoder that compresses video before sending was not released after the camera was turned off. After the fix, the same scenario took 30.
3 MB. The story is useful to anyone planning to write code together with a neural network: AI delivers the skeleton and the interface fast, but memory leaks, audio quality and security holes you have to hunt down yourself.
Why another chat app
The author just wanted to play with friends. He needed voice rooms, text chat, screen sharing, a decent noise suppressor and a client that doesn't take resources away from the game. He was tired of Discord's ads and Nitro offers, and without workarounds it doesn't even open for him. In TeamSpeak the noise suppression performed poorly, and Mumble lacked screen sharing. AI wrote the core of Nimbo, while the human handled the logic, the interface, the choice of stack and the architecture.
He first built a Windows client in C++ and Qt: C++ gives precise control over memory, and his friends have old laptops. But after every change he had to wait for an EXE build and hand the file out to his friends. By the author's estimate, design and UX work on that cycle burned three times as many tokens. So he froze the client and moved to the web: Vue and TypeScript, a Go server, and the ready-made LiveKit for transmitting voice, camera and screen.
A trick helped with the design. Off-the-shelf UI kits kept running into their own limitations, and a blank sheet made it hard to explain to the neural network what he wanted. The author collected references for mood and colors and gave them to the AI instead of a framework. The first screen came together quickly, but the model got confused by details such as spacing, corner rounding, borders and dividers, so there were many corrections.
Noise suppression: mouse clicks and the crunch of an apple
He started with RNNoise: it is lightweight and adds almost no latency. But in gaming sessions every mouse click and the rustle of a bag of nuts were audible. For comparison, he took ready-made recordings of mouse clicks, biting an apple and chewing it, ran them through three filters and measured how many decibels of the interference each one suppressed. The bigger the number, the better.
On three mouse recordings, RNNoise scored 13.55, 15.19 and 28.82 dB, GTCRN 19.21, 21.13 and 27.66 dB, and Speex 0.33, 0.93 and 1.61 dB. On the apple, the difference became huge: RNNoise 0.35 dB, GTCRN 21.29 dB, Speex 0.17 dB. The price of GTCRN is a slightly duller voice and a processing latency of 32 ms versus 21 ms for RNNoise. In a game this is not noticeable, so GTCRN became the default mode and RNNoise remained as an alternative. Processing runs on the user's computer, and optimizing that remains a question for the future.
Where AI didn't help: memory and security
On memory, the author had to become the tester: open a chat, watch a video, switch the camera on and off, and track after which click consumption failed to return to its baseline. The AI did not think through these scenarios on its own, although the author concedes that he may simply have prompted badly. For videos in chat, the client now frees the decoder on pause and remembers the position. As a result, a voice call on the test bench took 19.9 MB of physical memory across Nimbo's processes, and in real-world conditions you should expect 25–30 MB.
The author checked security like this: he ran some of the models locally, told the agents about the relevant piece of the architecture, formed a specific hypothesis for a bypass and sent them to test it on a test bench. That is how they found a vulnerability: an anonymous logout from the service could completely overwrite the database. The author is sure there will be more findings, and they are already sitting in the backlog.
The bottom line
There is no shared user database: the nickname, settings and server access keys live in a .nimbo file, which is protected by a mandatory password and opens in any compatible client. The code is open under the MIT license: the server, the web client and the Windows client are on GitHub. The downside is a real one: a server owner will have to rent a VPS and figure out deployment. Planned features include drawing over the screen, a game overlay, launching the server directly from the app, and clients for Mac and Android.
The practical takeaway: when vibe coding, budget time for measurements up front. AI will not check memory, audio or security on its own.
- Nov 21, 2026A viral GTA 6 Claude clip looks fake: half an hour of control burned through the 5-hour Pro limit
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
- Nov 21, 2026Waymo pulls robotaxis off highways after new entries into work zones
- Nov 21, 2026Google Opens Pre-Orders for Gemini Smart Glasses from $395, Shipping December 10
Key Questions Answered
Why another chat app?
The author just wanted to play with friends. He needed voice rooms, text chat, screen sharing, a decent noise suppressor and a client that doesn't take resources away from the game. He was tired of Discord's ads and Nitro offers, and without workarounds it doesn't even open for him. In TeamSpeak the noise suppression performed poorly, and Mumble lacked screen sharing. AI wrote the core of Nimbo, while the human handled the logic, the interface, the choice of stack and the architecture.
What are the key details on Noise suppression: mouse clicks and the crunch of an apple?
He started with RNNoise: it is lightweight and adds almost no latency. But in gaming sessions every mouse click and the rustle of a bag of nuts were audible. For comparison, he took ready-made recordings of mouse clicks, biting an apple and chewing it, ran them through three filters and measured how many decibels of the interference each one suppressed. The bigger the number, the better.
Where AI didn't help: memory and security?
On memory, the author had to become the tester: open a chat, watch a video, switch the camera on and off, and track after which click consumption failed to return to its baseline. The AI did not think through these scenarios on its own, although the author concedes that he may simply have prompted badly. For videos in chat, the client now frees the decoder on pause and remembers the position.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.