your coding agent wastes half its tokens just figuring out how your project works
OpenWiki from LangChain fixes that
it writes documentation for your repo that agents read as memory
then updates it every day from new commits and opens a PR
humans get an interactive map of the code to explore
personal mode turns your Notion, Gmail and web research into your own wiki
works with Claude Code, Codex, Cursor and OpenCode
free and open source, MIT license
your coding agent wastes half its tokens just figuring out how your project works
OpenWiki from LangChain fixes that
it writes documentation for your repo that agents read as memory
then updates it every day from new commits and opens a PR
humans get an interactive map of the code to explore
personal mode turns your Notion, Gmail and web research into your own wiki
works with Claude Code, Codex, Cursor and OpenCode
free and open source, MIT license
1 in 4 public AI agent skills has a security hole, according to research NVIDIA cites
so NVIDIA built a scanner to check them before you install
it's called SkillSpector
point it at any skill: a GitHub repo, URL, zip or folder
it reads the code and the instructions without running anything
catches hidden prompt injections, leaked secrets, data theft and poisoned MCP tools
plugs into CI with standard security reports
works for Claude Code, Codex, Gemini CLI and Cursor skills
free and open source, Apache 2.0, a "safe" result still needs a human look
1 in 4 public AI agent skills has a security hole, according to research NVIDIA cites
so NVIDIA built a scanner to check them before you install
it's called SkillSpector
point it at any skill: a GitHub repo, URL, zip or folder
it reads the code and the instructions without running anything
catches hidden prompt injections, leaked secrets, data theft and poisoned MCP tools
plugs into CI with standard security reports
works for Claude Code, Codex, Gemini CLI and Cursor skills
free and open source, Apache 2.0, a "safe" result still needs a human look
1 in 4 public AI agent skills has a security hole, according to research NVIDIA cites
so NVIDIA built a scanner to check them before you install
it's called SkillSpector
point it at any skill: a GitHub repo, URL, zip or folder
it reads the code and the instructions without running anything
catches hidden prompt injections, leaked secrets, data theft and poisoned MCP tools
plugs into CI with standard security reports
works for Claude Code, Codex, Gemini CLI and Cursor skills
free and open source, Apache 2.0, a "safe" result still needs a human look
someone built an AI that scored 740+ job offers for him and landed him a Head of Applied AI role
now it's free for everyone
it's called career-ops
paste a job link, it scores the offer against your CV from 1 to 5
below 4, it tells you not to bother applying
writes a tailored CV and cover letter for each role
scans company career pages and tracks every application on a dashboard
you still press submit yourself
works in Claude Code, Codex, OpenCode and Copilot, MIT license
someone built an AI that scored 740+ job offers for him and landed him a Head of Applied AI role
now it's free for everyone
it's called career-ops
paste a job link, it scores the offer against your CV from 1 to 5
below 4, it tells you not to bother applying
writes a tailored CV and cover letter for each role
scans company career pages and tracks every application on a dashboard
you still press submit yourself
works in Claude Code, Codex, OpenCode and Copilot, MIT license
someone built an AI that scored 740+ job offers for him and landed him a Head of Applied AI role
now it's free for everyone
it's called career-ops
paste a job link, it scores the offer against your CV from 1 to 5
below 4, it tells you not to bother applying
writes a tailored CV and cover letter for each role
scans company career pages and tracks every application on a dashboard
you still press submit yourself
works in Claude Code, Codex, OpenCode and Copilot, MIT license
free open source tools that replace your AI subscriptions:
1. YuE2 → instead of Suno, songs with vocals
2. VoiceStudio → instead of ElevenLabs, voices in 646 languages
3. LTX-2 → instead of Sora, 4K video with sound
4. PhotoCraft → instead of Photoshop, opens real PSD files
5. Continue → instead of GitHub Copilot, private coding help
6. Open WebUI → instead of ChatGPT Plus, your own chat app
check each license before using them for business
save this before your next renewal 💵
free open source tools that replace your AI subscriptions:
1. YuE2 → instead of Suno, songs with vocals
2. VoiceStudio → instead of ElevenLabs, voices in 646 languages
3. LTX-2 → instead of Sora, 4K video with sound
4. PhotoCraft → instead of Photoshop, opens real PSD files
5. Continue → instead of GitHub Copilot, private coding help
6. Open WebUI → instead of ChatGPT Plus, your own chat app
check each license before using them for business
save this before your next renewal 💵
free open source tools that replace your AI subscriptions:
1. YuE2 → instead of Suno, songs with vocals
2. VoiceStudio → instead of ElevenLabs, voices in 646 languages
3. LTX-2 → instead of Sora, 4K video with sound
4. PhotoCraft → instead of Photoshop, opens real PSD files
5. Continue → instead of GitHub Copilot, private coding help
6. Open WebUI → instead of ChatGPT Plus, your own chat app
check each license before using them for business
save this before your next renewal 💵
open AI models are catching up fast
the best open models you can download right now:
1. Kimi K3.1 Pro → ~3T parameters, close to Claude Opus 5.5 on Moonshot's tests
2. DeepSeek V4.1 Flash → MIT license, very cheap to run
3. Mistral Large 4 → strongest western open model at coding
4. Meta Muse Spark Open → built for long agent tasks
5. Ripple-65B → small, Apache 2.0, strong at terminal coding
read each license before you build a business on it
save this 🔖
open AI models are catching up fast
the best open models you can download right now:
1. Kimi K3.1 Pro → ~3T parameters, close to Claude Opus 5.5 on Moonshot's tests
2. DeepSeek V4.1 Flash → MIT license, very cheap to run
3. Mistral Large 4 → strongest western open model at coding
4. Meta Muse Spark Open → built for long agent tasks
5. Ripple-65B → small, Apache 2.0, strong at terminal coding
read each license before you build a business on it
save this 🔖
open AI models are catching up fast
the best open models you can download right now:
1. Kimi K3.1 Pro → ~3T parameters, close to Claude Opus 5.5 on Moonshot's tests
2. DeepSeek V4.1 Flash → MIT license, very cheap to run
3. Mistral Large 4 → strongest western open model at coding
4. Meta Muse Spark Open → built for long agent tasks
5. Ripple-65B → small, Apache 2.0, strong at terminal coding
read each license before you build a business on it
save this 🔖
the engine behind most self-hosted AI apps
it's called vLLM
it serves AI models much faster by managing GPU memory smartly
handles many users at once without slowing to a crawl
gives you an OpenAI-compatible API, so your apps don't need changes
runs Qwen, DeepSeek, Gemma, Llama, gpt-oss and more
used under the hood by tools like GPUWorks
free and open source
the engine behind most self-hosted AI apps
it's called vLLM
it serves AI models much faster by managing GPU memory smartly
handles many users at once without slowing to a crawl
gives you an OpenAI-compatible API, so your apps don't need changes
runs Qwen, DeepSeek, Gemma, Llama, gpt-oss and more
used under the hood by tools like GPUWorks
free and open source
the engine behind most self-hosted AI apps
it's called vLLM
it serves AI models much faster by managing GPU memory smartly
handles many users at once without slowing to a crawl
gives you an OpenAI-compatible API, so your apps don't need changes
runs Qwen, DeepSeek, Gemma, Llama, gpt-oss and more
used under the hood by tools like GPUWorks
free and open source
your phone can now run a full AI coding agent
it's called DSH Go
it puts DeepSeek Harness on Android, no root needed
run the agent right on your phone, or control the one on your computer from anywhere
it can even read your phone screen and tap for you, with payments and passwords always left to you
voice input, images, plugins
built by one solo developer, 18,000+ stars on GitHub in about a week
free and open source, AGPL-3.0
your phone can now run a full AI coding agent
it's called DSH Go
it puts DeepSeek Harness on Android, no root needed
run the agent right on your phone, or control the one on your computer from anywhere
it can even read your phone screen and tap for you, with payments and passwords always left to you
voice input, images, plugins
built by one solo developer, 18,000+ stars on GitHub in about a week
free and open source, AGPL-3.0
your phone can now run a full AI coding agent
it's called DSH Go
it puts DeepSeek Harness on Android, no root needed
run the agent right on your phone, or control the one on your computer from anywhere
it can even read your phone screen and tap for you, with payments and passwords always left to you
voice input, images, plugins
built by one solo developer, 18,000+ stars on GitHub in about a week
free and open source, AGPL-3.0
IBM's 8B model matches its old 32B model
it's called Granite 4.1
3B, 8B and 30B sizes
the 8B beats IBM's previous 32B model on tool calling, coding and following instructions
up to 512K tokens of context
no long "thinking", so answers come fast and cheap
also includes a speech model with a 5.33% error rate and a safety guard model
on Ollama and Hugging Face, Apache 2.0, free for business
IBM's 8B model matches its old 32B model
it's called Granite 4.1
3B, 8B and 30B sizes
the 8B beats IBM's previous 32B model on tool calling, coding and following instructions
up to 512K tokens of context
no long "thinking", so answers come fast and cheap
also includes a speech model with a 5.33% error rate and a safety guard model
on Ollama and Hugging Face, Apache 2.0, free for business
IBM's 8B model matches its old 32B model
it's called Granite 4.1
3B, 8B and 30B sizes
the 8B beats IBM's previous 32B model on tool calling, coding and following instructions
up to 512K tokens of context
no long "thinking", so answers come fast and cheap
also includes a speech model with a 5.33% error rate and a safety guard model
on Ollama and Hugging Face, Apache 2.0, free for business
6 free tools to keep your AI agents from becoming a security hole:
1. Codex Security → OpenAI's scanner that finds and verifies bugs in your code
2. garak → NVIDIA's tool that attacks your chatbot like a hacker would
3. OpenShell → NVIDIA's sandbox that asks before an agent touches anything risky
4. Strands Box → AWS's sandbox built so agents can't escape it
5. IronClaw → a personal assistant that never sees your passwords
6. PentAGI → AI agents that pentest your own systems
all open source
save this before you give an agent access to anything important 🔖
6 free tools to keep your AI agents from becoming a security hole:
1. Codex Security → OpenAI's scanner that finds and verifies bugs in your code
2. garak → NVIDIA's tool that attacks your chatbot like a hacker would
3. OpenShell → NVIDIA's sandbox that asks before an agent touches anything risky
4. Strands Box → AWS's sandbox built so agents can't escape it
5. IronClaw → a personal assistant that never sees your passwords
6. PentAGI → AI agents that pentest your own systems
all open source
save this before you give an agent access to anything important 🔖
6 free tools to keep your AI agents from becoming a security hole:
1. Codex Security → OpenAI's scanner that finds and verifies bugs in your code
2. garak → NVIDIA's tool that attacks your chatbot like a hacker would
3. OpenShell → NVIDIA's sandbox that asks before an agent touches anything risky
4. Strands Box → AWS's sandbox built so agents can't escape it
5. IronClaw → a personal assistant that never sees your passwords
6. PentAGI → AI agents that pentest your own systems
all open source
save this before you give an agent access to anything important 🔖
a free open image model that edits pictures using up to 8 reference images
it's called P-H3-Image, by Pruna AI
built on MiniMax's H3 video model, turned into an image generator and editor
keeps faces, products and details consistent between edits
up to 4K resolution
a 4MP image in 2.6 seconds on a B200, about 5x faster than the original, according to Pruna
works in ComfyUI
check the license: it excludes the US, EU, UK and South Korea unless you get a separate deal
a free open image model that edits pictures using up to 8 reference images
it's called P-H3-Image, by Pruna AI
built on MiniMax's H3 video model, turned into an image generator and editor
keeps faces, products and details consistent between edits
up to 4K resolution
a 4MP image in 2.6 seconds on a B200, about 5x faster than the original, according to Pruna
works in ComfyUI
check the license: it excludes the US, EU, UK and South Korea unless you get a separate deal
a free open image model that edits pictures using up to 8 reference images
it's called P-H3-Image, by Pruna AI
built on MiniMax's H3 video model, turned into an image generator and editor
keeps faces, products and details consistent between edits
up to 4K resolution
a 4MP image in 2.6 seconds on a B200, about 5x faster than the original, according to Pruna
works in ComfyUI
check the license: it excludes the US, EU, UK and South Korea unless you get a separate deal
Mistral built an AI that proves math is correct, and gave it away
it's called Leanstral 1.5
it writes proofs in Lean 4, a language where a computer checks every single step
solved 587 of 672 Putnam competition problems
about $4 per problem, vs an estimated $300+ for the closest rival
it also checks real code, and found 5 unknown bugs across 57 open source projects
only 6B parameters active at a time
free, Apache 2.0, plus a free API