Adrian Castellar

Practical AI for builders. I test new AI tools and models and show what works, what breaks, and what it costs. Cloud Engineer. AI-assisted, human-edited.

[Jun Song @jun_song]

Whatever new idea you want to build, most of it is already open source on GitHub.

Even if it is not, someone will clone it and open-source it within days of your launch.

The moat in coding is basically gone.

[Adrian Castellar @Cydr4l]

@RoryClear CPU only, inside a Docker container capped at 4 cores and 6 GB, no GPU passed through. docker stats showed it using about 1.4 of the 4 cores, so it didn't seem to parallelize. Is the tinygrad CPU backend mostly single-threaded? Happy to rerun it on a GPU.

[Adrian Castellar @Cydr4l]

clearcam adds object detection and tracking to any security camera, running locally with a small YOLO model.

I tested it on CPU in a container against a public tunnel camera. Setup took about 2.5 minutes, and the model download was only 8 MB. It ran at about 0.6 detections per second on 4 cores with 1.1 GB of RAM, and recording wrote about 5 MB per minute.

One catch: without a premium account it still sends alert notifications to https://t.co/WK64Qc6MAv by default. Change the notification URL in settings if you want it fully local.

Tip: install clang first. The README skips it, and it crashes without it.

https://t.co/9NBCEciDSr

[Adrian Castellar @Cydr4l]

Anthropic went from 2,500 merged changes a day in May to 30,000 in September, and git was the first thing to break.

The models get the headlines, but this is the number I'd watch. Most of our tooling still assumes a human pushes a few commits a day.

[Anthropic @AnthropicAI]

New on our Engineering blog: Creating in-house source control.

In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale.

Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: https://t.co/wzNkwHs7IN

[Adrian Castellar @Cydr4l]

Bruno is an open-source API client that saves your requests as plain text files, so your collections live in Git instead of a cloud account.

I tested the CLI on 4 requests against real APIs. The whole collection was 6 files and 1.7 KB, it ran in under 2 seconds, and one run wrote JSON, HTML and JUnit reports. Changing an assertion showed up as a one-line git diff.

One catch: I imported the Petstore OpenAPI spec and the first request returned a 400. The imported base URL was relative, and the query params never made it into the request.

https://t.co/RopVnzFceZ

[Adrian Castellar @Cydr4l]

Kandev is a kanban board for coding agents. Each task runs in its own git worktree, and you review the diff in the browser before anything ships.

I tested v0.97.0 in a container. One npx command installed it in 9 seconds, and the server was up in under a second. I plugged in a fake agent script that needs no API key, and it edited a file on its own branch while main stayed clean.

One catch: by default it listens on every network interface with no login, and its MCP server exposes 49 tools, including delete and settings changes. The server warns you about it. Turn on auth or keep it on localhost.

https://t.co/GKJ3PaEcCo

[Adrian Castellar @Cydr4l]

Anthropic-Cybersecurity-Skills is a library of 818 security skills for AI agents. It isn't made by Anthropic, the README says so.

I tested it. All 818 skills are real, and none of them are stubs. Then I ran the commands from two of them against the real tools. The Volatility3 Windows commands work, but the Linux one calls a plugin that doesn't exist. In the Sigma skill, 4 of 7 Python commands break on current pySigma.

Also, loading every skill's description into an agent costs about 70k tokens. Install only the domains you need.

https://t.co/oI2Nm2E0zJ

[Adrian Castellar @Cydr4l]

Ouroboros is an AI agent that rewrites its own code and keeps its memory across restarts.

I reviewed the repo before running it. The top contributor is the agent itself: an account called ouroboros-agent has 3,265 commits, more than 3 times its human author's 1,002.

It can run shell commands and scripts, use git and a browser, and start itself at sign-in. The docs say plainly that its tool layer is "not an OS sandbox", and an open issue reports that the Panic button misses sub-agent processes.

Interesting project. I'd only run it in a throwaway VM.

https://t.co/5JP4i1nvEr

[Adrian Castellar @Cydr4l]

f3d renders a 3D model to a PNG with one command, no GPU, no display, in about a second.

I tested it on 7 models, including an animated one. It's great for generating thumbnails in a pipeline.

Tip: plain OBJ files come out almost black unless you pass --color.

https://t.co/0G3d9S6tlH

[Adrian Castellar @Cydr4l]

slackdump saves your Slack history without needing admin rights.

I tested the offline part on a sample export. It's one 41 MB binary that installed in under 2 seconds, turned the export into a browsable HTML site in 49 ms, and ships an MCP server, so Claude can search your Slack.

One catch: the viewer shows mentions as raw user IDs.

https://t.co/InAe616WRv

[Adrian Castellar @Cydr4l]

Yesterday I reviewed designlang and it picked neon yellow as Linear's brand color.

So I dug in. It was ranking any bright color as "brand", and a big card link counted as a button. I sent a fix: Linear now comes out indigo, Stripe's text color is right, and all 687 tests pass.

https://t.co/5tyIevSe7R

[Adrian Castellar @Cydr4l]

@jun_song The real danger is falling behind because you didn't use it.

[Adrian Castellar @Cydr4l]

@jun_song On times like these you either ride with the wave or you get swallowed by it. But it won't last forever.

[Adrian Castellar @Cydr4l]

ODS seems to be trending. One installer that sets up about 20 local AI services, including llama.cpp, Open WebUI, n8n and ComfyUI.

Before you run it, know that it published 5 security advisories this week. The critical one: the admin API was reachable without signing in.

The fixes are only on main, not in the v3.0.0 release, and the install script pulls main anyway.

Keep it on localhost.

https://t.co/xXBM2cnddK

[Adrian Castellar @Cydr4l]

I ran designlang on Stripe and Linear. Each site took 9 to 14 seconds and produced 35+ files, including an MCP server for Claude Code.

It got Stripe's branding right and Linear's completely wrong.

A useful starting point, but not a source of truth.

https://t.co/TTNCFY1wBn

[Adrian Castellar @Cydr4l]

@Teknium This is awesome, going to try this one definitely

[Adrian Castellar @Cydr4l]

@theo rough translation: I learned the word webmcp today

[Adrian Castellar @Cydr4l]

@BenjaminDEKR one of these is going to age a lot better than the other

[Adrian Castellar @Cydr4l]

@mr_r0b0t same reason I paste screenshots into claude code all the time

[Adrian Castellar @Cydr4l]

@mr_r0b0t @NousResearch the gap between what you think your agent does and what the logs say is wild