IMURUN v0.3.0 · MIT · Windows, macOS, Linux

AI that lives
on your machine.

Local-first AI chat with its own inference engine. One process, no cloud: FISH, our C++ engine, runs GGUF models on the CPU; a Python server hosts sessions, tools and connectors; the Liquid Paper UI opens in a native window or your browser.

Install IMURUN Read the 23 guides

Install

One command, then just imurun.

Everything IMURUN writes — sessions, models, logs, settings — goes to ~/.imurun. Nothing is written into site-packages, and nothing is sent anywhere.

Python 3.11–3.14. Wheels for Windows x64, Linux x64 and macOS arm64 include the FISH engine.

Inside IMURUN

Seven parts, one window.

FISH our own inference engine

Engine · C++

Memory-maps a GGUF, keeps weights quantised, quantises activations to int8 once per matmul and runs AVX2 integer dot products. A KV prefix cache makes repeated system prompts free.

Liquid Paper no build step

Interface

Plain HTML, CSS and JavaScript, served under a strict CSP — no CDN, no remote font, nothing leaves. A native window, or your browser.

Tools that ask before they act

Agents

The model calls tools itself in a turn loop; anything that writes shows an approval card first. Skills and sub-agents included.

Web search with no API key

Research

The model decides whether to search and writes the queries; what reaches it is cleaned, ranked passages with numbered sources.

Eight accounts and any MCP server

Connectors

Gmail, GitHub, Slack, Vercel, Render, Cloudflare, Supabase and Canva — OAuth tokens in your OS keychain.

Speech, locally both directions

Voice

Whisper for speech-to-text and Piper for text-to-speech, running on the machine instead of a vendor's cloud.

Gets better from your own use

Tuning · opt-in

Thumbs-up the answers you like, teach the ones that were wrong; the tuning channel trains a LoRA on this machine, checks it did not get worse, and offers a new GGUF with a Use it button. Off by default.

How a turn works

From prompt to answer.

Should this turn search?

One short routing generation with the same local model (at most 60 tokens) decides whether the web is needed and writes the queries. Greetings and code pastes skip it entirely.

The research planner
2.1×int8 kernel vs the f32 path, per Q4_K row
5architecture families: llama, qwen2, qwen3, phi3, gemma2
8connected accounts, plus any MCP server
0bytes sent to a cloud by default

Kernel figure from the engine benchmarks (Ryzen 7 7435HS, Llama-3.2-1B Q4_K_M) — read the caveats.

Llama 3.xMistralQwen 2.5Qwen 3Phi-3Gemma 2DeepSeek-R1 distillsMixtralTinyLlamaSmolLMLlama 3.xMistralQwen 2.5Qwen 3Phi-3Gemma 2DeepSeek-R1 distillsMixtralTinyLlamaSmolLM

Security

Private by construction.

Local is not a setting you have to find. It is how the app is built.

  • Loopback onlyThe server binds to your machine, with a Host-header check.
  • Per-run tokenA bearer token injected into the page — never in URLs or logs.
  • Strict CSPdefault-src 'self': no CDN, no remote fonts or images.
  • KeychainOAuth tokens live in the OS keychain, not in a file.
  • Approval cardsEvery write action waits for you.
  • 1 MB capRequest bodies are capped; nothing runs unbounded.

Run it

Flags, and where the engine comes from.

The pip wheels for three platforms ship the engine inside. Elsewhere the app starts and tells you how to get one.

The full setup guide →
--window
native window (pywebview), the default when installed
--browser
open in your browser instead
--no-open
start the server only
--port N
choose the port
--model path.gguf
start with a model on disk
--get-engine
download the FISH engine for this platform, checksum-verified
PlatformEngineBuilt with
Windows x64fish.dllMinGW-w64 GCC, static
Linux x64libfish.soglibc 2.28+
macOS arm64libfish.dylibApple clang
Everything else—imurun --get-engine or a source build

Documentation · 23 guides

Start with a guide

All 23 guides
Start hereGuide 01 · Start here · 2 min read

Getting started

Local-first AI chat with its own inference engine. One process, no cloud: a C++ engine (FISH) runs GGUF models on the CPU, a Python server hosts sessions, tools and connectors, and a vanilla-JS "Liquid Paper" UI opens in a native…

Read the guide
  1. Start here · 7 min readInstall & runInstall IMURUN with pip or npx, what each package contains, where your data lives, the launch flags, and how to get or build the FISH engine on any platform.
  2. FISH engine · 7 min readFISH engine architectureFISH is a self-contained CPU inference engine for GGUF models: its own GGUF reader, quantised kernels, tokenizer, sampler, KV cache and hardware planner. It links nothing from llama.cpp or any other runtime. It is compiled to one shared…
  3. Agents & research · 9 min readAgents: tools, permissions, skillsHow the assistant calls tools by itself: the turn loop, what each tool may touch, approval cards for anything that writes, and skills and sub-agents.
  4. Connectors · 5 min readConnectorsIMURUN can act on eight external accounts — Gmail, GitHub, Slack, Vercel, Render, Cloudflare, Supabase and Canva — through tools named <service>_<verb> in the agentic tool loop (docs/agentic/README.md), and through MCP (MCP.md).

Have a model on disk?
Let's give it a window.

Install IMURUN Read the source ↗support@imutechz.in