DEDUPEDWhat people actually made with the repos we cover: things they built, benchmarks they ran and setups they explained. We read each post on X and sum up what was built, how and the result. The chips take you to the repo and to a use case you can try yourself, with steps and a copy-paste agent prompt; each card links to the original post.
-
Built with
Cut 60 GB of talking-head footage into 6 rough-cut videos with OpenClaw and Claude Code
On a dedicated Ubuntu mini PC, OpenClaw driving Claude Code ingested 40 minutes of footage, copied the creator's DaVinci folder structure, split dual-mic audio into two tracks, cut silences and mistakes, generated captions and exported a .drb timeline per video, ready to finish in DaVinci on a Mac.
OpenClawSee 3 ways to use OpenClaw
By @Naaackers
-
Built with
Built an AI fighting game intro video with HyperFrames and Claude Code
No After Effects or timeline: HyperFrames turned HTML into every slam, flash and health bar, media-use pulled in the sound effects and hyperframes-cli checked each frame and rendered the final video. Character art came from Nano Banana Pro.
HyperframesTry it: Make a product launch video from your website with an AI agent
By Rey Brando (@ReyBrando1)
-
Built with
Ran local LLMs on openSUSE with Ollama and Open WebUI in Podman
The author set up Ollama on an openSUSE Linux box with the web UI in a Podman container for local LLM chats.
OllamaTry it: Run an open model on your machine in one command
By furiyuri (@StevenLHess)
-
Hands-on
Compared VoiceStudio with ElevenLabs on real projects and wired it into Claude Code
The author installed VoiceStudio next to ElevenLabs: it handles most everyday voice work well but still trails ElevenLabs on long content and nuanced voices. Its MCP support connects cleanly to Claude Code and other agents.
Voice StudioSee 6 ways to use Voice Studio
By Serhiy Krytskyi (@kritzkidj)
-
Built with
Automated renewal reminders and follow-up calls with a custom Supabase CRM
For an insurance agency, the author built a CRM on Supabase that automates renewal reminders and follow-up calls.
SupabaseSee 5 ways to use Supabase
By Felix Lopez (@felixlopmsc)
-
Built with
Prototyped agents that answer with tables, forms and approve buttons on OpenUI
The author built a small OpenUI prototype with mock data for three business flows to test agents that reply with UI instead of paragraphs.
OpenuiSee 5 ways to use Openui
By Abhishek Ramkrishna (@abhi_ramkrishna)
-
Built with
Built an interactive 3D diagram of a user profile request with the Anatomy skill
With Anatomy, the author modeled the path terminal, gateway, server, cache, database as a rotatable machine that shows a cache hit, a 401 on an expired token and a 404 for an unknown user.
AnatomySee 4 ways to use Anatomy
By Muhammad Ahmad (@muhammad_ahmaad)
-
Benchmark
Wrote custom kernels that beat SGLang by 1.154x on an NVIDIA H200
In the Open Kernel Challenge the author passed all 9 correctness tests at 1.154x SGLang's speed, cutting router time from 10.2 to 3.9 microseconds with a custom Triton kernel. TMA (36.5 us) and Split-K (50.1 us) attempts were slower than SGLang's 25 us.
SGLangSee 3 ways to use SGLang
By 𝐕𝐚𝐥𝐞𝐫𝐢𝐢 (@valeriixai)
-
Built with
Built a Grok bot that answers US law questions on Firecrawl's free law API
Firecrawl published 30M+ pages of US law behind a free API with no key, and the author built a Grok bot on top of it.
FirecrawlSee 3 ways to use Firecrawl
By Luke Chase (@metalfox)
-
Built with
Built an fzf floating pane on top of Herdr to track blocked and working agents
The author runs the Herdr TUI in Ghostty and wrote a wrapper that queries Herdr and uses fzf to show a floating, tabbable list of blocked, working and all agents across OpenCode and Pi.
HerdrTry it: Run several coding agents and see which one is stuck
By Mark Jaquith (@markjaquith)
-
Hands-on
Tested OpenSpec 1.15.0's update dry run and found it removes local edits
The author customized a generated Claude command, previewed the update with --dry-run --json (no file hashes changed), then applied it and watched the local note disappear. The write-up includes a disposable Python check you can rerun before refreshing your assistant's instructions.
OpenSpecSee 5 ways to use OpenSpec
By Shivam Pandit (@Pandit12Shivam)
-
Built with
Replaced a TV's home screen with an ad-free launcher using Codex and Free My TV
The author gave Codex the Free My TV repo, the TV's IP and region; it handled the setup, and a customized launcher gave a clean ad-free interface without buying a new device.
Free My TVSee 4 ways to use Free My TV
By Rohit Kanojiya (@rohitkanojiya73)
-
Hands-on
Turned old PCs into servers and moved all work off a MacBook with T3 Code
The author connects to repurposed laptops and PCs through T3 Code from laptop and phone, and has moved coworkers onto it since the orchestration update.
T3 CodeSee 3 ways to use T3 Code
By Jaden Maciel-Shapiro (@JadenLShapiro)
-
Built with
Built a personal AI assistant on Supabase
The author built their own AI assistant with Supabase as the backend; janky, they say, but it has been working well.
SupabaseSee 5 ways to use Supabase
By Unsupervised Ape (@unsupervisedape)
-
Benchmark
Cut Supabase local rebuilds from 142 to 21 seconds by running it without Docker
The author tried Supabase's native runtime on a codebase with 1,192 migrations: 21 seconds vs 142 on an M3 Max and 56 vs 248 on an Ubuntu VM. The write-up covers what broke along the way, from initdb refusing root to branch switches that pick a different local database.
SupabaseSee 5 ways to use Supabase
By Ariel (@rajmaliuk)
-
Built with
Built a chat app on LiteLLM with a PHP server and deployed it on Railway
The author used LiteLLM as the model gateway behind a PHP backend, shipped the app on Railway and had it certified with eMudhra.
LitellmSee 5 ways to use Litellm
By samagra14 (@samagra_sharma)
-
Benchmark
Runs DeepSeek 1017 in production on Project Maya at about 50 tok/s
The author's new production setup serves DeepSeek 1017 locally with Project Maya, measuring about 50 tok/s decode and 1,400 to 1,600 tok/s prefill.
Project MayaSee 5 ways to use Project Maya
By mgabor (@mgabor_)
-
Hands-on
Tried PhotoCraft as a free Photoshop replacement for family use
Many basic features work and some do not; frame animation, which the author does with the kids, does not work yet, but it could save paying for 6 licenses.
PhotoCraftSee 3 ways to use PhotoCraft
By Zak (@zearl)
-
Built with
Designed snap-on knight armor for 6-inch figures with a SolidCraft CAD bot
The author built Caliper, a Grok bot that models in SolidCraft, and used it to make printable armor that snaps onto a child's poseable figures.
SolidcraftSee 6 ways to use Solidcraft
By Raymond Thurman (@raythurman2386)
-
Benchmark
Measured Repomix compression: 33% fewer tokens, but constants can vanish
With repomix 1.18.1, --compress plus --remove-comments cut a demo from 869 to 580 tokens, and a 618-file repo packed to 1.37M tokens in 5.4 seconds. Caveats found: compression dropped a constant, a secret in src/config.js was not flagged, and a second run packs its own output file.
RepomixSee 3 ways to use Repomix
By @ReignSwords
-
Built with
Turned a post-it sketch into a parametric 3D model with SOMA
The author first modeled the part with a self-published FreeCAD plugin, then redid it in SOMA, the CadQuery desktop app built for coding agents, and found the agent workflow and lighter UI much nicer.
By John Kattenhorn (@johnjkattenhorn)
-
Hands-on
Tried SolidCraft as a Fusion 360 style CAD app on Linux
The author found it snappy and liked that UI elements can be resized.
SolidcraftSee 6 ways to use Solidcraft
By Mark Ping (@markping)
-
Benchmark
Ran a 125B LLM at about 40 tok/s on a mini PC plus eGPU with llama.cpp
On a Ryzen AI Max+ 395 mini PC (128GB) with a Radeon AI PRO R9700 over USB4, the author split Qwen3.8-Flash-Next 125B across both AMD GPUs with llama.cpp and ROCm on Windows, hitting 37.7 tok/s sampled and 41.2 tok/s greedy. Getting two AMD architectures to cooperate took a lot of debugging.
llama.cppSee 3 ways to use llama.cpp
By Yuxuan Cao (@ycaophysics)
-
Hands-on
Used AI agents to speed up herdr-gpui and got 4 PRs merged upstream
With 50+ agents over SSH the app lagged, so the author pointed Sol and Opus 5.5 at the repo to find the slowness, then followed the contribution guidelines and landed 4 merged PRs.
Herdr GpuiSee 4 ways to use Herdr Gpui
By Miguel Ríos Berríos (@MiguelriosEN)
-
Benchmark
Made a Claude coding harness 39% faster with Harness2
The author adapted Harness2 to optimize their own harness; after the first optimization run Claude ran 39% faster and scored 10/10 on accuracy for their codebase's tasks.
Harness2See 7 ways to use Harness2
By Daniel Müller (@BrahmaD111)
-
Built with
Self-hosted Rakazo on a private VPS to have agents handle business expenses
The author runs Rakazo on their own VPS, where agents with their own computers sort out the expenses for the author's business.
RakazoSee 4 ways to use Rakazo
By Cyris (@SudoLabsNZ)
-
Benchmark
Benchmarked lithos-metal against Ollama on a daily Qwen3.8-27B voice workload
On an M5 Pro, lithos-metal was meaningfully faster than Ollama after warmup for Voxt, the author's local voice IME (about 5.6K-token system prompt, short replies). Ollama still wins on serving features like keep_alive, idle unload and near-zero idle CPU.
Lithos MetalSee 4 ways to use Lithos Metal
By Guitar Cat + LLM 「Will @ SGLang Summit in SF」 (@GenAI_is_real)
-
Hands-on
Ran the CodeAF coding agent fully local on Ubuntu with Qwen3.8-27B on Ollama
The author installed CodeAF, a single Go binary, and watched its planner, worker and checker models split a job into tasks, run them and test the results, with per-task cost by model. The verdict: an early preview with rough edges, shown in a full video walkthrough.
CodeAFTry it: Drive a coding agent on a remote dev box from your laptop
By Fahd Mirza (@fahdmirza)
-
Hands-on
Upgraded agents from plain browsing to screenshot reports with agent-browser
After switching to Vercel's agent-browser, the author's agents now produce reports with screenshots instead of just surfing pages.
Agent BrowserTry it: Take annotated or skip-if-unchanged screenshots for agent checks
By ZhenAI (@zhenyaThinks)
-
Built with
One-shotted VEED's launch video with GPT-6 ASTRA on OpenEdit
The author gave the model only the website and brand guidelines, and OpenEdit rendered the full launch video in one pass.
Open EditSee 5 ways to use Open Edit
By Sabba Keynejad (@sab8a)
-
Hands-on
Ran a local coding agent on Windows with Unsloth Desktop and Qwen3.8-27B
After failing to get Hermes Agent working on Windows, the author installed Unsloth Desktop with no friction and uses its Project mode (plans, browses, writes code) with a local Qwen3.8-27B GGUF, so no quota runs out.
UnslothTry it: Run Claude Code on a local quantized model
By @mauno_k
-
Hands-on
Got Gemini 3.8 Flash to finish a full plan in one prompt using the SureForge skill
With the Claude quota used up, the author gave Gemini 3.8 Flash a detailed plan plus the SureForge skill and it completed the whole task on the first try with no errors. Without the skill in a later prompt, the results dropped right back.
Sure ForgeSee 4 ways to use Sure Forge
By NDQ (@NDQ_VN)
-
Built with
Built a working sheet metal louver cassette with Codex and text-to-cad
A mechanical engineer had text-to-cad generate the parametric build123d source, STEP assembly, previews, OSH Cut DXFs and an order list, with about 13 major tweaks from first prompt to ordering. Lessons: inspect the STEP before DXFs, use section views, and upload DXFs early since OSH Cut caught invalid folds.
Text to CadSee 4 ways to use Text to Cad
By Ian Lawson (@IanL556)
-
Benchmark
Ran a 125B model at 25 tok/s on one RTX 4090 with Unsloth's MTP support
With Unsloth's new MTP support, the author ran Qwen3.8-Flash-Next (125B total, 6B active) on a single 24GB RTX 4090 at 80K context and got about 25 tok/s.
UnslothSee 3 ways to use Unsloth
By lemon👑 (@lemonDefi1)
-
Built with
Built Chattr, a dashboard around the last30days skill to browse past research runs
The author wrapped last30days-skill in a dashboard called Chattr that captures the chatter around a topic across platforms and keeps past runs easy to look up.
Last30days SkillSee 3 ways to use Last30days Skill
By Fred Moon 🚀 (@fw3d)
-
Built with
Routed Claude Code through FreeLLMAPI across Mistral, Gemini, Groq and Cerebras
The author set up FreeLLMAPI behind Claude Code, added four free providers, tested balanced routing and automatic fallback, and dug into provider failures and rate limits.
FreellmapiSee 5 ways to use Freellmapi
By Yellow void (@yellow_void22)
-
Benchmark
Benchmarked KittenTTS against Kokoro for local voice agents on CPU and GPU
On a shared Colab CPU the 15M KittenTTS Nano model made 24 seconds of audio in 15 seconds (1.6x real time), while forcing ONNX onto a T4 GPU was only 1.1x faster than the CPU. The author shared the Colab code and the three model sizes tested (15M, 40M, 80M).
Kitten TTSSee 3 ways to use Kitten TTS
By Alok (@analogalok)
-
Built with
Made an explainer video for a side project with OpenMontage and Claude
The author had OpenMontage and Claude produce a video explaining SingleLog, a personal project, in a few minutes and at a low cost because the project is small.
OpenMontageSee 5 ways to use OpenMontage
By Gordon Murray (@gortron)
-
Built with
3D printed a modular cube kit designed with text-to-cad
The author vibe coded a set of geometric modular cubes with text-to-cad, printed them, and plans to use the workflow for more personal projects.
Text to CadSee 4 ways to use Text to Cad
By Yogi Seezatnap (@seezatnap)
-
Built with
Designed a 3D printed erector-style building set for kids with text-to-cad
The author used text-to-cad with a 3D printer to prototype a toy construction set the kids now play with, and is considering SendCutSend for metal parts next.
Text to CadSee 4 ways to use Text to Cad
By Jordan Hall (@jordanhall)
-
Benchmark
Grew a site from 22 to 58 SEO pages in one session with the claude-seo plugin
Using the claude-seo plugin in Claude Code, the author added competitor comparison and language pages, built a changelog from git history, fixed a canonical URL bug and submitted everything through Google's indexing API. The takeaway: with one backlink and 3 impressions in 28 days, links mattered more than content.
Claude SEOTry it: Run a full SEO audit on your site from Claude Code
By KylianM (@leonlykiks)
-
Built with
Updated a PR comment triage skill to use the GitHub MCP server's new reply tool
The author switched a review-pr-comments skill to the GitHub MCP server's new tool for replying to inline PR comments, making it faster to triage review comments together with Copilot.
GitHub MCP ServerSee 3 ways to use GitHub MCP Server
By Pamela Fox (@pamelafox)
-
Built with
Built a YouTube competitor analysis dashboard in under 30 minutes with Ruflo
Using Ruflo on Claude Code, the author built a dashboard that shows top videos, best titles, rising topics and working content angles for any keyword, then rewrites scripts in their own voice.
RufloTry it: Add the Ruflo console to Claude Code
By Stinson_edits | Video Editor (@stinson_edits)
-
Benchmark
Generated a CAD practice model from just a YouTube link with Codex and text-to-cad
Given only the video link, Codex read the drawing, built the model with the text-to-cad workflow, exported CAD files, rendered review images, estimated the mass and made an exploded view in under 30 minutes. Some geometry was still wrong, but it needed almost no intervention.
Text to CadTry it: Generate a 3D CAD model from a text description
By 許文彥 (@vm3jp6u04)
-
Hands-on
Replaced half of a custom UI codebase with OpenUI v0.5
The AlphaWeb team has used OpenUI since launch; v0.5's built-in data layer, incremental editing and inline mode replaced much of the code they had written themselves.
OpenuiSee 5 ways to use Openui
By Anurag Sharan (@anuragsharan_)
-
Benchmark
Built a sub-300 ms voice pipeline with PersonaPlex and Cartesia Sonic Turbo
The author pairs NVIDIA's PersonaPlex speech model with Cartesia Sonic Turbo and a custom dialog script enforcer to keep voice replies under about 300 ms of latency.
PersonaPlexTry it: Talk to a full duplex voice agent in your browser
By GPTina (@chatGPTina)
-
Built with
Built a Shopify data sync with Claude Code using Context7 for API docs
The author had Claude read both platforms' API docs through Context7, plan and build the sync, then fixed mismatches by feeding it real API objects, with Claude writing a new test for each case.
Context7Try it: Stop your agent coding against outdated library APIs
By Philo Hermans (@Philo01)
-
Hands-on
Tried LangExtract on FAIR principle examples and its visual output
The author wrote simple FAIR principle examples for Google's LangExtract, ran it on existing text and found the visual output useful and the extraction decent. Next step: a BAML version to compare outputs.
LangExtractSee 3 ways to use LangExtract
By @fils
-
Built with
Built a Firecrawl alternative that turns docs into LLM-ready Markdown with crawl4ai
The author used crawl4ai to convert any documentation site into Markdown files ready to use as LLM context.
Crawl4AITry it: Turn any web page into clean Markdown for RAG
By juggernaut (@curlysaarthak)
Builds are found by searching X for each repo. We keep a post only when it links to or names that exact repo and shows someone building with it, benchmarking it or explaining how to use it, and only when we can say clearly what was built; complaints and questions are left out. Headlines and summaries are written by Deduped from the post and its links; each post belongs to its author. Deduped is not affiliated with the authors or the repos. Want a post removed? Reply to any Deduped email.
Every Monday: the 5 hottest repos and exactly how to use them. Free, one email a week, unsubscribe anytime.
Get them free