Top Dev Tool Projects : Spec Kit, RAGFlow, Unsloth, T3 Code, Paritok & OpenViking Transcript, AI Summary & Key Points

00:00 Every week, developers release new open-source dev tools that make building software faster and easier. This is your weekly DevTool project update video. Top trending open-source and best DevTool projects this week. Without wasting time, let's get started. >> Before we jump into today's project updates, here's a quick announcement for everyone. We've launched a brand new YouTube channel called AI Agent Studio dedicated entirely to AI Agent projects, tutorials, and tools.

00:27 So, if you're interested in staying up to date with the latest AI agent open source projects, learning how to build your own agents, or exploring cuttingedge agent frameworks, make sure to check it out. Subscribe now to get weekly videos, in-depth guides, and real time project breakdowns. The link is right there in the description. Don't miss it. All right, let's get into today's video.

00:52 >> Project number one, SpecKit. Specdriven development toolkit for AI agents. SpecKit is an open- source toolkit from GitHub for building software with AI coding agents through a specd-driven process. Instead of prompting an agent to write code straight away, you make the specification the source of truth and let it drive the implementation. You install a command line tool called specify, initialize a project for your agent, and work through a set of /comands.

01:17 One writes your project's guiding principles. One defines what to build and why. One turns that into a technical plan for your chosen stack. One breaks the plan into tasks and one executes them. Optional commands clarify gaps, check consistency, and generate quality checklists. It works with more than 30 coding agents across CLIs and IDEs, and you extend or reshape it with extensions, presets, and role-based bundles.

01:42 It runs on Linux, Mac OS, and Windows with Python and Git. It suits developers and teams who want structure around AI built software. Install the CLI and write your first spec. Project number two, Ragflow. Open-source Rag engine with agent capabilities. Ragflow is an open-source retrieval augmented generation engine that builds a context layer between your documents and large language models.

02:06 It centers on deep document understanding. So it parses messy real world files like Word, Slides, Excel, plain text, images, scanned copies, structured data, and web pages. Then chunks them with explainable template-based rules you can inspect and adjust by hand. From that, it retrieves relevant passages using multiple recall paths with fused ranking and answers with grounded traceable citations back to the source to reduce hallucination.

02:34 It adds an agentic layer with pre-built templates, memory, MCP support, and connectors to sources like Confluence, S3, Notion, and Google Drive. You plug in your own chat and embedding models, self-host it with Docker Compose, or run it from source and integrate it through clean APIs. It suits developers and enterprises building question answering over their own data.

02:55 Deploy it and turn your documents into a grounded assistant. Project number three, Unsloth. run and fine-tune AI models locally. Unsloth is an open- source tool for running, training, and deploying AI models on your own machine. It covers language, diffusion, embedding, audio, and vision models and lets you fine-tune them faster with far less memory through custom Triton kernels, so training that normally needs a large GPU fits on modest hardware.

03:23 It comes in three forms, a native desktop app, a web UI called studio and a code-based core you install with pip. It supports Laura Qura full fine-tuning, pre-training and reinforcement learning methods. Builds data sets from PDFs, CSVs, and DOCX files and exports to formats like GGUF for deployment. You can serve models through an open AI or anthropic compatible API.

03:48 connect them to coding agents like Claude Code and Codeex as local backends and reach them remotely through a secure tunnel. It runs on CPU, Nvidia, AMD, Intel, and Apple hardware, including multiGPU setups. It suits developers and researchers customizing open models, install it, and fine-tune your first model. Project number four, D3 code control coding agents from any device.

04:14 D3 Code is an open-source control surface for the AI coding agents already running on your machine. Rather than being an agent itself, it drives the ones you have through a mobile app, a web app, and an Electron desktop app. So, you can start and steer coding work from your phone or browser instead of sitting at the terminal. It works with your existing subscriptions to Claude Code, Codeex, Cursor, Grock, Build, and Open Code.

04:38 Once a provider is installed and logged in on your computer, T3 code can control it. You run its backend locally with a single command which also serves the local web app and it is built to be remote ready so you can reach your agents from anywhere. The project is deliberately open so you can fork it if it ever moves in a direction you dislike. It suits developers who run agents and want to manage them away from their desk.

05:02 Run it and control your agents from your phone. Project number five, Reverse Skill. Skill Router for AI security work. Reverse Skill is an open-source skill router pack that helps AI coding agents handle reverse engineering, authorized penetration testing, and security research. When an agent meets an APK, a binary encrypted front-end JavaScript, a CTF challenge, or a pentest target, it stops the agent from guessing commands and instead routes the task to the right methodology, checks which tools are present, and runs

05:32 a repeatable evidence-based workflow. A single structured routing configuration maps tasks to scenario skills covering Android and iOS binaries.NET malware and Yara firmware exploit development API security and more with a dedicated orchestrator holding dozens of CTF subsklls. It enforces a scope gate so an agent confirms authorization before acting.

05:56 connects to real tools like Judag, IDA Pro, Raider 2, Freda, and Burp Suite through scripts and MCP servers and keeps a case timeline and field journal. It stays client neutral working with clawed code, codecs, cursor, client and others, and runs on Windows, Linux, Mac OS, and Kali. It suits security researchers who work with AI agents. Clone it and route your first target.

06:21 Project number six, X for you feed algorithm. Open- source code behind X's feed. X for you feed algorithm is the open- source code that decides which posts a viewer sees in the for you feed on X. It assembles the feed per request from two pools. Recent posts from accounts the viewer follows held in memory and post from accounts they don't follow. Found through a machine learning retrieval model and cluster similarity search.

06:46 A transformer model called Phoenix reads the viewer's recent activity and predicts for each post how likely they are to take each action on it. And those predictions are combined into one score with weights held openly in the code. Ranking then applies author diversity decay, an outof network discount, and a new author boost before a separate service reorders for variety.

07:07 A distinct visibility filtering path attaches safety labels to posts and accounts and decides whether each post is shown, hidden, or placed behind an interstitial. It ships runnable pieces including training and synthetic data for the model and a transparency tool showing labels on your account. It suits developers and researchers auditing recommendation systems.

07:26 Read the code and see how your feed is built. Project number seven, H3 Metal Native Miniax H3 video engine for Macs. H3 Metal is an open-source native inference engine for running the Miniax H3 video model on Apple Silicon Max. Written in C with metal GPU kernels, it is a focused self-contained engine for one model family rather than a general runner.

07:48 It turns a text prompt into a short video with synchronized audio and also supports first and last frame conditioning, an ordered image, video, and audio references so you can anchor or continue a scene. You point it at a downloaded model snapshot and run it from the command line or an interactive session that keeps the model in memory between prompts.

08:09 It exposes fine control over the speed quality trade-off letting you tune dnoising passes, active transformer layers, internal render resolution and duration, and it can preview frames directly in a graphical terminal. An SSD streaming mode reads model blocks from disk, so it runs on machines with limited memory, and it uses ffmpeg to encode the final MP4.

08:29 It suits developers and creators, generating video locally on a Mac. Build it and render your first clip. Project number eight, Parat context compression model for coding agents. Parat is an open-source language model trained specifically to compress the context that AI coding agents send to an LLM. It is a small 4 billion parameter Laura adapter over QN34B trained on real coding agent trajectories.

08:57 So, it can tell a function signature or file path from a throwaway debug line, keeping what matters and dropping what doesn't. It runs as a middleware proxy between your agent and the model API. It intercepts each request, compresses tool results and old history, and trims the tool schemas down to the ones in use, then forwards the smaller prompt upstream, so you are build on fewer input tokens.

09:18 Compression is never destructive since trimmed content is tagged with a reference the model can expand on demand through a virtual tool. You self-host it on a single 24 GB GPU through Alma or VLLM or point it at a hosted endpoint and your agent just changes its base URL. It works with clawed code, cursor, codecs and other standard format agents. It suits developers running long agent sessions who want to cut token costs.

09:47 Start the proxy and route your agent through it. Project number nine, Open Viking context database for AI agents. Open Viking is an open-source context database for AI agents that unifies their memory, knowledge, and skills in one place. Instead of scattering context across vector stores, it organizes everything as a virtual file stem under a Viking my/ipath.

10:08 So, an agent locates and browses information with familiar commands like ls, find, and grep rather than opaque semantic lookups. Each entry is stored in three tiers. A oneline abstract, an overview, and the full details loaded on demand to keep token use low. Retrieval works recursively, first locating the most relevant directory, then drilling into it layer by layer, and the whole path is recorded so you can see exactly why something was returned.

10:36 At the end of a session, it can extract long-term user preferences and reusable task experience, so the agent improves with use. You self-host it with your own embedding and vision models through providers like OpenAI, Olama, or others. And it integrates with agents such as Claude Code, Codeex, and Cursor. It suits developers building agents that need durable memory and knowledge.

10:58 Deploy it and give your agent a persistent brain. Project number 10, Skills for Real Engineers, agent skills for disciplined software work. Skills for real engineers is an open-source collection of agent skills for building real applications with coding agents. Rather than vibe coding, the author packages his own daily use skills as small composable pieces you can adapt, each targeting a common failure mode to fix misalignment.

11:23 Grilling skills make the agent interview you in depth before it writes anything. And a companion skill builds a shared project vocabulary in a context. MD, so both sides speak the same language. For quality, a test-driven development skill drives a red green refactor loop. A bug diagnosis skill runs a disciplined phase gated debugging cycle, and a code review skill checks the diff against both standards and the original spec.

11:48 Others turn a conversation into a spec, break work into tickets, plan large multisession efforts, and survey a codebase for architectural improvements. Skills are either user invoke to orchestrate or model invoke to hold reusable discipline. You install them as a claude code plugin or copy editable files into your project. For many agents, it suits engineers who want their agent to follow real practices.

12:11 Install the skills and start with a grilling session. Project number 11, OMLX menu bar. LLM inference server for Macs. OMLX is an open-source LLM inference server built for Apple Silicon and managed from the Mac OS menu bar. It serves text models, vision language models, OCR models, embeddings, and rerankers through one open AI and anthropic compatible endpoint, so any standard client can connect.

12:37 Its standout feature is a two-tier KV cache. Frequently used blocks stay in RAM, and when memory fills, blocks spill to SSD and are restored on a matching request instead of being recomputed, surviving even a server restart. Continuous batching handles concurrent requests, and a multimodel engine loads several models at once with least recently used eviction, pinning, and per model idle timeouts to stay within a memory ceiling.

13:04 A web admin dashboard covers monitoring, downloading models from hugging face, benchmarking, and one-click setup for coding agents like Codeex and Copilot, plus a built-in chat. A native Swift menu bar app starts and monitors the server without a terminal. It runs on M1 through M4 Macs and installs as an app through Homebrew or from source. It suits Mac users running local models for real work.

13:30 Install it and serve a model from your menu bar. Project number 12, Node. Node client for the Amadeus protocol network. Node is the open- source node software for the Amdus protocol, a peer-to-peer network you join by running its Damon built in Elixir and distributed as a self-contained binary called Amodusd. It connects to the network over UDP, exposes an RPC API, and can run unattended as a systemd service with auto updates and crash restart.

13:55 It supports web assembly smart contracts. The repository ships assembly script samples, and from an interactive session, you can deploy a compiled contract and call its functions alongside built-in operations like transferring the network's coin. A local testn net mode lets you run the whole thing on one machine and point a browser at a local RPC endpoint to experiment.

14:18 Nodes can take on roles through environment flags running as a plane participant, a block producer or a validator. It targets Linux tested on recent ibuntu and builds through docker or podman. The maintainers describe it as an experimental research project shared for educational purposes. It suits developers who want to run or study the network, build it and start a local test net.

14:39 Project number 13, public APS community curated directory of free APIs. Public Apes is an open-source communitymaintained directory of free public APIs for building software. It is a single reference list rather than a tool or library. Developers browse it to find an API for whatever they're building from weather and finance to games, machine learning, and open data.

15:02 Entries are grouped into categories and presented in tables. And each one notes practical details a developer needs before integrating, such as whether it requires an API key or other authentication, whether it serves requests over HTTPS and its core support. Everything lives in a markdown readme you read directly on GitHub. And the project grows through pull requests with contribution guidelines and automated link checks keeping listings consistent and current.

15:29 It is one of the most widely referenced resources of its kind and is maintained with backing from a player. It suits developers and learners looking for data sources or side project ideas. Open the list and find an API for your next project. Project number 14, Nell LM Fit. Find which LLMs run well on your hardware. LLM fit is an open-source terminal tool that right sizes local language models to your machine.

15:53 It detects your CPU, RAM, and GPU, then scores hundreds of models across quality, speed, memory fit, and context, and tells you which ones will actually run well rather than just fit. Written in Rust, it ships an interactive terminal UI with a classic command line mode. And for each model, it picks the best quantization that fits, estimates tokens per second from your hardware's memory bandwidth, and handles multiGPU and mixture of expert setups where only some experts load at once.

16:22 A plan mode inverts the question, estimating what hardware a given model config needs. And a simulation mode lets you test other machines. It reads which models you already have from run times like a Lama, Llama, CPP, MLX, LM Studio, and Docker Model Runner, downloads new ones, benchmarks live, and can pull crowdsourced realworld speeds from other users.

16:46 It runs on Linux, Mac OS, and Windows, and installs many ways. It suits anyone running local models. Install it and see what fits your machine. Project number 15. Dashen Freebuff. Free ADUP supported terminal coding agent. Freebuff is an open-source free coding agent that edits your codebase from the terminal through natural language instructions. It is the adupported edition of codebuff.

17:09 So there is no subscription, no credits, and no configuration. You install it, open your project, and tell it what you want. And it finds the right files, makes the changes, and runs your tests. Rather than using one model for everything, it coordinates specialized agents, one to scan the codebase and pick files, one to plan the changes, one to make precise edits, and one to review them, which gives better context and fewer errors.

17:38 It runs on strong open-source models like DeepSeek, MIMO, and Miniaax with no proprietary lockin and adds built-in web research and browser use. Features include file and agent mentions, a bash mode, resumable chat history, and per project knowledge files. It offers a full mode in some countries, and a limited mode elsewhere, and is funded by text ads, so it analyzes prompts to personalize them.

18:02 It suits developers who want a capable coding agent for free. Install it and start coding in your terminal. Project number 16. Wiser. One ownership rule for memory threads networks. Wiser is an open-source staticallytyped compiled programming language built around a single idea. Many hard bugs. Whether memory errors, deadlocks, or network failures come from unclear ownership of a resource.

18:26 So Wiser uses one ownership rule to govern all three. It aims for the safety of Rust without the steep learning curve. managing memory through Perseus reference counting, an approach borrowed from Koka and Lean 4 that gives seike speed without a garbage collector or a borrow checker for concurrency and networking. It uses choreographic programming where you write one description of an interaction and the compiler generates and checks the code for every participant catching deadlocks and dropped messages before the

18:55 program runs. The same ownership rule extends to threads and hardware interrupts. The syntax is deliberately plain with immutable by default variables, explicit types, strrus, standard control flow, and errors returned as result values rather than hidden exceptions. The maintainers are candid that it is early research with unsolved problems. It suits language enthusiasts and researchers interested in distributed safety.

19:21 Read the docs and try writing a program. Project number 17, mast anti-slop rules to stop AI generated generic UI. Anti-slop is an open-source rules document that stops AI coding agents from producing generic templated AI slop interfaces. It centers on a single markdown file. The agent reads on demand only when a task involves UI or UX work rather than being loaded into every session.

19:48 The file list recognizable slot patterns to avoid such as blue purple gradients, overused glass morphism, and marketing buzzwords. then lays out three dozen mandatory rules and a set of preference agnostic craftsmanship criteria covering intentionality, functional completeness, and evidence over claims. It is a filter rather than a style guide. So, it imposes no specific colors, fonts, or layouts and leaves brand direction to you.

20:14 Before delivering, the agent must run a checklist that pairs one verification question with each rule and report the results. You wire it in with a single pointer line in your existing agent instructions file which keeps it tool agnostic across cloud code, codecs, cursor and others and portable across projects. It suits developers who build UI with coding agents.

20:35 Add the file and point your agent at it. Project number 18, Vocat web control panel for cellular modems. Vocat is an open-source web control panel and engineering toolkit for Qutel EC20 and EC25class cellular modems. It gathers many modem tasks that normally need separate tools into one self-contained service you reach through a browser. It discovers attached USB and serial modems automatically.

21:00 Shows live radio status like operator, signal strength, band, and registration and gives you interactive AT and USSD terminals with full command history. You send and receive SMS over both cellular and IMS. Set up Wi-Fi calling through an IIV 2 tunnel with IMS registration and manage ESIM profiles on supported cards listing switching downloading or deleting them.

21:25 It also handles network selection sasops proxy routing per device and forwards incoming messages through telegram email or web hooks with a telegram bot for remote control. The backend is written in Go with a React and TypeScript front end embedded into a single binary that stores state in SQLite. It runs on Linux across several architectures, installs through a script or docker and self-updates with checksum verification.

21:54 It suits people working hands-on with cellular hardware. Install it and open the panel for your modem. Project number 19, mange airship Figma- like visual editor for real code bases. Airship is a command line tool that gives developers a visual editor for their existing web app without adding plugins or dependencies. It runs as a reverse proxy in front of a dev server showing the live app on an infinite canvas where desktop and mobile frames sit side by side.

22:21 A person clicks an element describes the change and a connected coding agent cla codeex or open code edits the actual source file. Everything runs locally using credentials already set up for the chosen agent. This suits front-end developers who want design level control without leaving their codebase. Point it at your dev server and start editing visually.

22:40 Project number 20. MCTIRS. Open-source template for AI powered directory sites. MCTIRS is an open-source website template that helps developers build directory websites. The kind that lists businesses, tools, or resources under categories and tags. It handles listings, search, user accounts, and paid or sponsored submissions through Stripe. While AI assists with content generation during submission, the template runs on Nex.js and React, uses sanity as a headless CMS for content and blog posts, OJ JS for

23:12 authentication, and resend for email and newsletters. It deploys to Versel or Docker. This suits developers who want a ready-made foundation instead of building directory features from scratch. Clone the repository and start customizing your own directory site. Thanks for watching. See you in the next update.