Weekly · open-source AI · week 43, 2026
scout + 13 more trending open-source AI repos · week 43, 2026
This week's trending open-source AI on GitHub — scout, blender-ai-mcp-addon, indra, and more: the newest, Hacker News talk, a subfield spotlight, and the fastest-rising, drawn out of the noise.
Newest this week
Scout: the open-source browser agent, by Outcrawl. Mirrored automatically from Outcrawl's own repository.
TypeScript · 946 stars already
View on GitHub →What we said about scout
This one hands an AI agent the keys to an actual browser, so it can click, scroll, fill forms, and pull data the way a person would instead of relying on brittle API scraping. It's called Scout, built by the Outcrawl team and mirrored straight from their own repo. What stands out is that it's written in TypeScript, which means if you're already living in the Node and Playwright world, you can read the source, patch it, and bend it to your workflow without switching mental models. Browser agents are crowded territory right now — you've got Browser Use on the Python side and a handful of closed SaaS tools — so an open, TypeScript-native option fills a real gap for frontend and full-stack folks. It's aimed at people automating repetitive web tasks, running research pipelines, or testing flows that don't expose clean APIs. Nearly a thousand stars this early tells me developers are hungry for something they can self-host and audit. Worth cloning if you want transparency over a black box.
AI assistant and MCP server for Blender: visual chat, scene tools, multi-agent workflows. Works with OpenAI, ChatGPT, Claude, Gemini, Ollama, OpenCode. Blender AI addon and plugin, copilot for 3D, bpy Python scripting, MCP integration.
Python · 790 stars already
View on GitHub →What we said about blender-ai-mcp-addon
This one lets you talk to Blender in plain language and have an AI actually manipulate your scene — adjusting objects, writing bpy scripts, running multi-step edits through a chat panel right inside the viewport. It's called blender-ai-mcp-addon, and what makes it interesting is the MCP server underneath. Because it speaks the Model Context Protocol, you're not locked into one provider — OpenAI, Claude, Gemini, even a local Ollama model all plug in the same way, so you can keep sensitive project work on your own machine if you want. That flexibility is the real differentiator here. A lot of Blender AI tools hardwire a single API and leave you stuck when pricing or policies change. This treats the model as swappable infrastructure instead. It's aimed at technical artists and TD-leaning folks who already live in Python but would rather describe an operation than hunt for the right bpy call. Nearly eight hundred stars this early suggests it's scratching a genuine itch. If you're curious about agentic workflows in a real creative tool, this is a grounded place to experiment.
Fast CPU OCR: line detection, CTC recognition, reading order, markdown
Python · 638 stars already
View on GitHub →What we said about indra
This one pulls structured text out of documents without ever touching a GPU, running line detection, recognition, reading order, and markdown export entirely on your CPU. It's called indra, from the datalab team, and the interesting part here is the deliberate CPU-first design. Most OCR stacks you reach for—think the heavyweight transformer-based models—assume you've got a graphics card sitting around, which makes them awkward for cheap servers, edge devices, or just batch jobs where you don't want to pay for accelerated compute. Indra leans on CTC recognition, which is lighter and faster than autoregressive decoding, and it handles reading order so your output doesn't come back as scrambled columns. That markdown export is the thoughtful touch—you get something you can feed straight into a RAG pipeline or a docs system without extra cleanup. If you're processing invoices, scanned PDFs, or forms at volume and latency matters more than squeezing out that last fraction of accuracy, this is worth benchmarking against Tesseract or PaddleOCR. Six hundred stars this early suggests plenty of people share that CPU constraint.
Talk of Hacker News
Lightweight coding agent that runs in your terminal
Rust · 170 points on HN
View on GitHub →What we said about codex
Imagine a coding assistant that lives right where you already work—your terminal—reading your files, running commands, and iterating on changes without you ever opening a browser tab. That's the pitch behind openai/codex, a lightweight agent written in Rust that keeps the whole loop local and fast. The Rust choice matters here: you get a single snappy binary, low memory overhead, and none of the cold-start lag you feel with heavier Node or Python wrappers. Where tools like Aider or Continue lean into editor integration, Codex bets on the command line being the real control center, which clicks nicely if you live in tmux and vim. It's aimed at developers who want an agent they can script, pipe, and automate rather than click through. The honest caveat is that terminal agents still ask for trust—they run things on your machine—so you'll want to understand its sandboxing and approval flow before turning it loose on a repo. If your workflow is keyboard-first, this is worth a serious afternoon of experimentation.
A read-only MCP server that lets your AI agent search and read your own WhatsApp messages. Runs on Cloudflare Workers.
TypeScript · 85 points on HN
View on GitHub →What we said about whatsapp-mcp
Picture asking your AI assistant to pull up that address a friend sent you on WhatsApp three weeks ago, and getting it back in seconds. That's the pitch behind whatsapp-mcp, a read-only MCP server that gives your agent search access to your own WhatsApp history. The key word there is read-only — it can look, but it can't send, which is a sensible guardrail when you're wiring a language model into your private conversations. What makes this one practical is that it runs on Cloudflare Workers, so you're not babysitting a server or exposing a desktop client to the internet. Compared to the handful of WhatsApp bridges floating around that demand full send-and-receive permissions, this trades capability for a much smaller blast radius. It's aimed at developers already building personal agent workflows who want their chat history as a searchable source without handing over the keys. I'd still read the auth flow carefully before pointing anything at your account, but as a focused, deliberately limited tool, it's a thoughtful design.
Subfield spotlight
Open-source Next.js template for building apps that are fully generated by AI. By E2B.
TypeScript · 6,391 stars
View on GitHub →What we said about fragments
This one hands you a working skeleton for the kind of app where a user types a prompt and gets back a running piece of software, rendered live in the browser. It's called fragments, from the E2B team, and what makes it interesting is where the generated code actually executes. Instead of just spitting out a code block, it runs the AI's output inside a sandboxed cloud environment, so you see a real, interactive result — a chart, a small app, a data script — not just text pretending to be one. For developers, that sandbox piece is the hard part you'd normally spend weeks building, and here it's wired up alongside model support for Anthropic, OpenAI, Mistral and others. Compared to rolling your own with raw API calls, you skip the plumbing around secure execution and multi-provider switching. It's a solid starting point if you're prototyping an AI-builds-the-app experience and want something you can fork, read end to end, and actually understand rather than a closed black box.
A curated list of AI-powered coding tools
2,172 stars
View on GitHub →What we said about awesome-ai-coding-tools
This is a hand-maintained directory that catalogs the AI coding tools sprouting up across the ecosystem, from autocomplete engines to full agentic assistants. It's called awesome-ai-coding-tools, and while "awesome list" repos are a dime a dozen on GitHub, this one earns its place by sitting on top of a category that's changing almost weekly. The real value here isn't the list itself — it's the triage. Instead of chasing hype threads or scattered Reddit recommendations, you get one page grouped by what the tool actually does: code completion, review, testing, documentation. For a team lead evaluating what to pilot next quarter, that structure saves genuine hours. Compared to something like a standalone benchmark site, this doesn't rank or score — it surveys, which means you still do your own vetting. My honest take: treat it as a starting map, not a verdict. The maintenance cadence matters more than the star count, so before you trust it, check how recently the last pull request landed. A stale awesome list is just a museum.
Ship features 10x faster. Built In Auto: Memory, K8S Agent & Security (SDD+SDLC) .
TypeScript · 266 stars
View on GitHub →What we said about shep
This one bundles a handful of things developers usually stitch together by hand — persistent memory, a Kubernetes agent, and security checks woven into the development lifecycle. It's called shep, from shep-ai, and it leans into spec-driven development paired with the full SDLC rather than treating those as separate concerns. The pitch of shipping faster is familiar, but what's actually interesting here is the attempt to make the agent aware of your infrastructure and your security posture at the same time, instead of bolting those on after the fact. If you're already running on Kubernetes and tired of your AI tooling forgetting context between sessions, the memory piece alone might earn it a look. Compared to standalone coding agents, shep is aiming to be more of an opinionated pipeline than a chat window. It's early — a couple hundred stars, written in TypeScript — so I'd treat it as something to experiment with on a side project before trusting it near production. Worth watching how the K8S integration matures.
Fastest-rising
An agent skill for creating detailed, interactive isometric illustrations of real-world machines and complex systems. Generates polished SVGs with optional WebGL shaders and 3D effects.
JavaScript · +14,927 stars this week
View on GitHub →What we said about anatomy
This one hands your AI agent the ability to draw — specifically, to generate detailed isometric breakdowns of machines and complex systems as clean, interactive SVGs. It's called anatomy, from wheresryan22, and the thing that makes it click is the scope: not icons, not flowcharts, but exploded cutaway views of how real hardware fits together, with optional WebGL shaders if you want lighting and depth on top.
Where I see this landing is documentation and education. If you've ever tried to explain a pump assembly or a server rack layout and ended up screenshotting someone else's diagram, this is the alternative — generated from a prompt, editable because it's vector, not a flattened PNG.
Compared to piping text into a general image model, you get something you can actually crack open and tweak node by node. The tradeoff is it's opinionated toward that isometric style, so it's not your everyday charting library.
Worth a look if you're building technical docs, interactive manuals, or teaching tools and you're tired of diagrams nobody can maintain.
Editorial diagram design for Claude Code, Codex, GitHub Copilot, Cursor, Factory Droid, and Pi. 44 diagram types. Self-contained HTML + SVG. No shadows. No Mermaid slop.
HTML · +8,083 stars this week
View on GitHub →What we said about diagram-design
Here's a repo that gives your AI coding assistant a proper eye for diagrams, handing it a design system so the charts it generates actually look intentional. It's called diagram-design by Cathryn Lavery, and it's a set of editorial standards and templates that plug into Claude Code, Codex, Copilot, Cursor, and a few others. The idea is simple: instead of letting your agent spit out generic Mermaid output, you point it at this spec and get self-contained HTML and SVG across 44 diagram types — flowcharts, timelines, system architectures, the usual suspects, but with real typographic care.
What I appreciate is the opinionated restraint. No drop shadows, no clutter, just clean editorial layout you'd actually put in documentation or a deck. If you've ever watched an LLM produce a diagram that technically works but looks like it escaped from 2009, this is the correction. It's aimed at developers and technical writers who want their generated visuals to match the polish of their prose. Lightweight, framework-free, and refreshingly specific about taste.
Independent Auditing of AI Agents. Run by human or the agent itself, to answer the most crucial question in the AI Agent Economy. Is the agent doing what is supposed to do? With iFixAi you can have this answer in less than 120 seconds.
Python · +7,071 stars this week
View on GitHub →What we said about iFixAi
Picture an independent referee that watches your AI agent work and tells you, within about two minutes, whether it actually did the job it claimed. That's the pitch behind iFixAi, an auditing layer for autonomous agents that can be triggered by a human reviewer or by the agent checking its own behavior. The interesting part is the framing: as we hand agents real tasks — booking, buying, writing to databases — the gap between "the agent said it finished" and "the agent did the right thing" becomes a genuine liability. Most observability tools today show you traces and token counts; iFixAi is aiming higher, at outcome verification rather than plumbing. If you're shipping agents into production, this sits closer to a QA or compliance role than to logging. It's early, Python-based, and the self-auditing idea raises an obvious question — can an agent honestly grade itself? Worth watching, especially if you pair it with human spot-checks rather than trusting it blind.
Shipping hard
Orca is the place for working with a fleet of agents (e.g. Grok Bot, Claude Code). Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.
TypeScript · 3,195 commits this week
View on GitHub →What we said about orca
Imagine running a whole squad of coding agents at once, each doing its own thing, instead of babysitting one chat window at a time. That's the idea behind Orca from stablyai. It's a control layer for orchestrating multiple agents—think Claude Code, Grok Bot, whatever you've got—using your own subscriptions rather than renting someone else's API wrapper. What stands out to me is the reach: desktop, mobile, and a remote runtime, so you can kick off a task on your laptop and check progress from your phone on the train.
Where this differs from tools like OpenHands or Aider is the fleet mindset. Those tend to focus on one agent working a repo deeply. Orca leans into parallelism and coordination across many. That's most useful if you're juggling several branches or micro-tasks that don't need constant supervision.
The nearly thirty-two hundred commits this week tell you it's moving fast, which cuts both ways—expect rough edges. If you're comfortable managing agents at scale, it's worth a look.
SolidCraft: an open-source parametric 3D CAD application built in pure Rust. Desktop, web, and automation through MCP.
Rust · 618 commits this week
View on GitHub →What we said about solidcraft
Imagine designing a mechanical part where every dimension is a variable you can tweak, and the whole model rebuilds itself around that change. That's the parametric CAD workflow, and SolidCraft brings it to desktop and browser as an open-source project written entirely in Rust. What caught my eye here is that it ships with MCP support, meaning you can drive the modeling engine programmatically — think an agent or a script generating geometry instead of you clicking through menus. That's a genuinely different angle from the usual FreeCAD comparison, which leans heavily on Python macros and a decades-old codebase. The Rust foundation also hints at tighter memory behavior and a cleaner path to running the same geometry kernel in a WebAssembly build. Six hundred commits in a single week tells you this is early and moving fast, so temper your expectations — you're looking at a foundation, not a Fusion 360 replacement yet. But if you're a hardware hacker or a tools developer curious about scriptable, automatable CAD, this is worth cloning and watching closely.
Never stop coding. Free MIT AI gateway: one endpoint, 359 providers (150+ free), 1200+ models Kimi, Claude, GPT, Gemini, GLM, DeepSeek, MiniMax. Works with Claude Code, Codex, Cursor, OpenCode, Cline & Copilot. Quota-aware auto-fallback, RTK+Caveman compression saves 15-95% tokens, MCP/A2A, Desktop/PWA. Built by hundreds of contributors
TypeScript · 470 commits this week
View on GitHub →What we said about OmniRoute
Picture a single API endpoint that quietly routes your requests across hundreds of model providers, switching automatically when one hits its quota. That's OmniRoute, an MIT-licensed AI gateway in TypeScript that stitches together 359 providers and over a thousand models behind one interface. What I find genuinely useful here isn't the raw numbers — it's the quota-aware fallback. If your free tier on one provider dries up mid-session, it reroutes instead of failing, which is the kind of thing you only appreciate after a tool has broken your flow once. It plugs into Claude Code, Cursor, Cline, Copilot, and Codex, so you're not rewriting your setup. The token compression claims — 15 to 95 percent — are worth testing yourself, since savings depend heavily on your prompts. Compared to something like OpenRouter, the draw is self-hosting and that MIT license, meaning no middleman owns your routing layer. If you're juggling free tiers across providers and tired of managing keys by hand, this is worth an afternoon. Four hundred seventy commits this week, too — this project is moving fast.