shipped, in beta, in the works

Projects

Peridot, Madhuca, agent-wire and Cordon are shipped and still getting regular updates. Peridot 1.6.0 came out on 2 October and taught it to use tools. FluxBoard is in public beta. TuringOS is in active development. All of it is open source and worked on in public on the uncoalesced GitHub. parentheses-1.0-240M, a Kannada-first recurrent language model, came out in November with open weights.

Peridot

Shipped

A sovereign, local-first AI inference kernel. No cloud dependency, no telemetry, and you decide how the model runs.

Peridot runs large language models on hardware you own, fully offline, on Windows and Linux. Version 1.6.0 moved to a native CUDA engine: a 27B model now fits entirely on an 8 GB laptop GPU and decodes at 23.6 tokens a second. It can call skills and sandboxed plugins you approve, and Claude Code, Codex or Gemini CLI can hand it private jobs over MCP, and by default you review what goes back. Behavior is set by a versioned constitution file you can read and edit.

  • local inference
  • privacy-first
  • AI kernel
  • RAG
  • Windows + Linux
stars
last commit
loading live stats…Loading GitHub stats

Madhuca

Shipped

A free, open-source fire and smoke radar for India. Open the page, see where the fires are, and see which way their smoke is heading.

Pick a region and Madhuca pulls the latest NASA satellite fire detections, checks the wind at each fire, and draws it on a map of India. Each fire is tagged likely crop burning or likely wildfire, gets a smoke plume, and comes with a short alert in Hindi, Kannada, Telugu or English that can be read aloud. It runs on demand with no accounts and no background jobs, and a failed scan is shown as an error, never as an all-clear. Built with Aaron and Rahul for Google’s Build with AI: Code for Communities hackathon. MIT licensed.

  • wildfire and smoke
  • open source
  • Cloudflare Workers
  • MapLibre

TuringOS

In development

An agentic execution layer built on Debian. It runs Claude Code agents in Btrfs sandboxes, with a live desktop UI, an MCP tool registry and a Game Mode.

TuringOS puts a policy layer between Claude Code and your filesystem. Every agent task runs inside a throwaway Btrfs snapshot, and nothing touches the real project until a person has read the diff and approved the merge. It ships as a bootable Debian live ISO with a fullscreen desktop shell (a Brave app window with a Rust service behind it), an MCP tool registry called Clawd Bazaar, and a Game Mode that pushes agents to the back of the queue while you play.

  • agent tooling
  • sandboxing
  • Rust
  • Bash

agent-wire

Shipped

One MCP server that gives AI agents live access to the web and 31 platforms. Free, it runs on your machine, and the common cases need no API keys.

Claude, Codex, Cursor and the rest answer from training data that is months old, and their fetch tools stall on login walls, Cloudflare checks and raw HTML. agent-wire is one warm Python process that routes each request through a fetch waterfall, strips hidden text and escapes fake control tags on the way in, scrubs tokens and cookies on the way out, and hands back dense Markdown: a GitHub issue we measured went from ~198k tokens to ~2k. It also writes cited last-30-days research briefs across a dozen sources. MIT licensed, on PyPI as agent-wire-cli.

  • agent tooling
  • MCP
  • web access
  • Python

FluxBoard

Beta

A customisable Android keyboard with a sticker library attached, built so that nothing you type is sent to a server.

The keyboard glide-types, learns your words in a dictionary kept on the device (it only moves if you transfer it to another phone of yours), and can be re-themed, re-laid-out and resized. In password and PIN fields it stops learning, suggesting and capturing what you copy, and it works that out from the field itself rather than trusting the app to say so. Stickers are cut out of screenshots and photos by hand and stay editable, and a video clip can become a GIF or animated WebP. No accounts, no ads, no analytics, and CI fails the build if a Play Services or Firebase class gets in. Kotlin, Jetpack Compose, Android 8.0+.

  • Android
  • Kotlin
  • keyboard
  • privacy
  • open source

Cordon

Shipped

Resource control for AI coding agents, one tool call at a time. A pytest run and a git status need wildly different amounts of memory, and one container-wide limit can’t serve both.

Cordon hooks into Claude Code, Codex CLI, Gemini CLI, Cursor and Hermes, and measures memory and CPU for every tool call. Commands you run through it get their own cgroup v2 with a soft limit that slows them down instead of killing them, and it tells the agent when that happened. On macOS the limits are advisory. The part that would make those decisions inside the kernel is still waiting on kernel features that aren’t upstream yet, and the project says so plainly.

  • agent tooling
  • cgroup v2
  • resource limits
  • Python

A predictive preemption engine for local LLM inference, and part of Peridot. It watches typing and microphone activity at 10 Hz and frees the GPU just before a prompt arrives.

When a local model shares its GPU with a background job, every prompt cold-starts while the GPU sheds that work. ZAT-SCS estimates the chance that you are about to send a prompt and, once it crosses a threshold, reclaims GPU resources from background jobs and preloads your context before you press Enter. It is part of Peridot, not a separate product, and since Peridot 1.6.0 it is opt-in and off by default.

  • local inference
  • GPU scheduling
  • Python
  • part of Peridot

Verity

Verity reads invoices and receipts, checks each one against fourteen written accounts-payable controls, and says which are fine and which need a person.

Each document is read for supplier, date, amount, line items and a signature, compared with the last six months on file for duplicates and split invoices, and then cleared, held for review or blocked. Every finding quotes the exact policy wording it relies on. Verity never pays or approves anything: a named person decides on every hold, and an override is recorded as one. MIT licensed.

  • audit trail
  • document checks
  • Python

Argus Council

Three AI agents that research something for you and keep the receipts.

A Scout downloads sources, an Archivist logs each one in a ledger with its link, file hash, citation and exact quotes, and an Integrator writes the evidence into your paper, report or dataset. A small standard-library Python script checks every quote, hash and citation before the work can be called finished. It works with any agent that reads AGENTS.md and can run Python. MIT licensed.

  • agent tooling
  • research
  • provenance
  • Python

Chakravyuh

In development

An open-source, local-first voice agent that answers scam calls on behalf of the people they target.

It takes the call as a fictional, slightly confused elderly persona, keeps the caller talking, and records the payment and contact details they read out: UPI IDs, bank accounts with IFSC codes, phone numbers and links. A deterministic state machine decides what the persona is doing, a local model only words the line, and the result is a report a person can review and submit through official channels such as the National Cyber Crime Reporting Portal. Everything runs on your own machine. Built for the Next Gen Hackathon 2026 in Bengaluru (23 to 25 October).

  • voice agent
  • local-first
  • scam calls
  • Python

parentheses

Shipped

A recurrent language model for Kannada, Tamil, Telugu, Malayalam and English, built from scratch on one 8 GB laptop GPU. Codename Project Tanzanite.

The 1.0 release has 253.54M parameters and keeps 4.85 MB of state at any length, where the memory of a transformer grows with every token. Kannada is its first language. On nine public benchmarks it beats the peer baseline on four and trails it on four others. Apache 2.0 weights on Hugging Face, MIT code on GitHub.

  • language model
  • Kannada
  • open weights
  • Python