Open-source apps and tools for speech, language models and AI agents.
A meeting note taker for Linux that transcribes and summarises on the machine itself. iPhone and Mac apps built on Apple Intelligence, where the AI runs on the device. MCP servers that let coding agents message each other and render narrated videos. And servers and libraries for running Whisper and LLMs behind one OpenAI-compatible URL. All MIT-licensed, by Saurabh Sharma.
Open source
MeetingInsights for Linux
v0.1.6
Linux, GTK, faster-whisper, llama.cpp, on-device AI
Record a meeting on a Linux desktop and get structured minutes, with transcription and summarisation running on the machine itself.
GTK app plus a CLI. Your microphone goes to one audio channel and everything you hear to the other, so speaker labels come from routing rather than a diarization model guessing at voices. Decisions, questions and action items are extracted by rules; a small local model writes only the prose. Whisper via faster-whisper, llama.cpp bundled in the .deb, no GPU needed. The Linux counterpart of the MeetingInsights iPhone app.
curl -fsSL https://techbysaurabh.github.io/apt/install.sh | sh sudo apt install meetinginsightstask-broker-mcp
v0.1.0
MCP, agents, server-sent events, TypeScript
An MCP server that lets AI coding agents send messages to each other. One command, no database.
Claude Code, Cursor, Gemini CLI or any MCP client connects to one running server, registers under a name, and can message other agents directly or by broadcast. Delivery is pushed over MCP's streamable HTTP transport the moment a message arrives, with a per-agent replay buffer for anyone offline and a wait_for_message tool that blocks instead of polling. Consumption still depends on the receiving agent taking a turn, which the README says plainly.
npx github:techbysaurabh/task-broker-mcpvideo-mcp
v0.1.0
MCP, Remotion, Playwright, HyperFrames, ElevenLabs
An MCP server that turns a script into a narrated video: HTML pages, Remotion compositions or HyperFrames layouts, in 16:9 or 9:16.
Narration is generated first and its real duration drives the visuals, so every segment gets exactly the on-screen time its voice-over takes. Three engines share that timing model: headless Chromium scrolling an HTML page, Remotion rendering React templates or custom JSX, and HyperFrames rendering HTML plus GSAP styles. ffmpeg muxes the audio and writes captions from the same timestamps. HTTP and stdio transports; S3 upload optional.
git clone https://github.com/techbysaurabh/video-mcp && npm installllm-host
v0.2.8
vLLM, Whisper, Kokoro TTS, OpenAI API
OpenAI-compatible inference server: an LLM via vLLM, Whisper speech-to-text, and Kokoro text-to-speech at one URL.
One command brings up chat, transcription and speech endpoints that existing OpenAI clients can use by changing the base URL. Tuned to fit a Colab T4 by default, with larger models a flag away. A gateway Docker image plus compose file runs the full stack on any CUDA box.
pip install llm-hostwhisper-blaze
v0.1.18
CUDA, Hopper, Whisper, dynamic batching
High-throughput batched Whisper large-v3 serving, with VRAM capping and a fused Hopper LayerNorm kernel.
Concurrent requests are fused into a single generate pass instead of being decoded one chunk at a time. A VRAM_LIMIT_GB cap lets it share an 80 GB card with other work, and each request can choose a fast or accurate long-form mode. Ships as a pip package and an OpenAI-compatible Docker image.
pip install whisper-blaze --no-build-isolationvrambatch
v0.1.0
PyTorch, batching, OOM recovery
Run batched GPU work inside a VRAM budget, self-healing on CUDA out-of-memory.
Sizes each pass from a memory budget and a measured cost per item. When a pass still overflows, it is halved and retried down to a single item, so a bad estimate costs one extra pass rather than a failed request. Extracted from the whisper-blaze serving path; pure Python with torch optional.
pip install vrambatch
iPhone and Mac apps
Published on the App Store. All but DailySteps and TKPH Calc run their AI on the device with Apple Intelligence, so nothing you say or track leaves the device. Each app has its own page with features, support and a privacy policy.
MeetingInsights
On-device AI · iPhone · 2026
Records a meeting, transcribes it live, and writes the minutes on the phone.
Summary, decisions and action items with owners, generated by Apple Intelligence from an on-device transcript. Works offline, no account, nothing uploaded.
MeetingInsights for Mac
On-device AI · Mac · in App Review
Records both sides of a call and writes the minutes on your Mac.
Captures your microphone and the audio playing on the Mac, so the other side of a Zoom, Meet or Teams call is transcribed too. Apple Intelligence turns the transcript into a summary, decisions and action items. No cloud, no account, no subscription.
Wornwell
On-device AI · iPhone · in App Review
Finds the clothes you wear in your photos and builds your wardrobe on the phone.
Scans the photos you choose on the phone, with Apple's on-device Vision on iPhone and bundled open-source models on Android, makes a cut-out of every garment with its type and colours, and lets you review everything before it is added. On iPhone, Apple Intelligence adds descriptions, outfit ideas and plain-language search. Nothing is uploaded.
Moodly
On-device AI · iPhone · 2026
DailyStepsPro
On-device AI · iPhone · 2026
ChatAI
On-device AI · iPhone · 2026
DailySteps
No AI · iPhone · 2026
TKPH Calc
No AI · iPhone · in App Review
Checks a haul truck's tyres against their heat limit, offline, on the phone.
Computes site TKPH/TMPH per axle from the truck, the haul and the site temperature, applies the tyre maker's published correction method (Michelin-style, Bridgestone-style or Generic) and shows the margin and verdict. Saved studies, PDF and CSV reports, a what-if solver, a GPS cycle logger and a Hindi interface. No account, no server, nothing uploaded.
All apps by this developer: App Store developer page
About
Saurabh Sharma, Bengaluru
techbysaurabh is the identity I publish personal open-source work under: desktop and iPhone apps where the AI runs on the device, MCP servers for coding agents, and serving code for speech and language models.
Everything here is my own work, released under MIT. Issues and pull requests are welcome on GitHub. For anything else, write to contact@techbysaurabh.com. My career site is saurabh.page.