OFFLINE · OPEN SOURCE · ANY HARDWARE

Local AI for coding and security. On any computer.

litecode is an AI coding agent that runs on your own machine. It fixes bugs, adds features, runs tests and reviews security. It also writes, explains and works with data. It fits a 4 GB laptop with no GPU, and it scales up to a GPU workstation with 30B-class models.

Get started See the benchmarks

litecode · ryzen 3 3200g · 16 GB · no gpu

› fix the bug so python test_calc.py prints ok
  └ MiniCPM5-2B loaded
▸ read   calc.py
  └ 4 lines
▸ edit   calc.py
▲ edit a file?
    - return a - b
    + return a + b
  └ yes
▸ run    python test_calc.py
  └ exit 0 · ok
◆ The bug was in add(). It subtracted. I changed
  line 2 to return a + b. The test passes.
  96 tok · 14s · MiniCPM5-2B

What it does

Built for code and security. Useful for everything else.

The main focus is coding and security review. The same local models also write, explain and analyze, so one tool covers most daily work.

MAINCoding

Fix bugs across files, add features with tests, refactor safely, explain code, write commit messages. /test runs your tests and fixes failures.

MAINSecurity

Static scan with zero model tokens, then a 4-line verdict per finding. SQL injection, XSS, command injection, path traversal, secrets. How it works.

ALSOWriting

Emails, summaries, READMEs, API docs and reports in plain, short sentences that are easy to read in a second language.

ALSOStudy

Ask about concepts, learn a codebase, practice offline. No internet and no data charges after setup.

ALSOData

CSV reports, JSON config edits, log analysis. litecode writes the script and runs it to check the result.

ALSOYour tools

Connect MCP servers and save your own workflows as skills. Works with the common .mcp.json and SKILL.md formats. All use cases.

Measured, not claimed

Proven on a 2019 budget PC

We test on a weak machine on purpose: AMD Ryzen 3 3200G (4 cores), 16 GB DDR4-2133, no usable GPU, models on a hard drive. If litecode works here, it works better on yours.

Generation speed

15.2 tok/s

MiniCPM5-2B, CPU only

8B-class speed

13.2 tok/s

LFM2.5-8B-A1B (MoE), CPU only

Coding tasks

8/8 pass

MiniCPM5-2B, LFM2.5-8B-A1B, Gemma-4-E4B

Vulnerability classes

10/10 found

MiniCPM5-2B, Qwen3.5-4B, Gemma-4-E4B, Llama-3.2-3B

Security review

40 s

Scan + verdicts for 2 findings, incl. model load

Cloud calls

After setup, nothing leaves the machine

MEASURED llama.cpp CPU build, 4 threads. Full tables and 30 agent challenges: Benchmarks.

Any hardware

It fits the model to the machine

litecode reads your RAM, CPU, disk and GPU, then picks models that fit. Small machines get small, fast models. Big machines and GPUs get bigger, stronger ones.

4 GB

system

Qwen2.5-Coder-1.5B

fim 0.5B

8 GB

system

MiniCPM5-2B

fim 1.5B

16 GB

system

LFM2.5-8B-A1B

fim

32 GB

system

LFM2.5-8B

Qwen3.5-9B quality

fim

GPU 24 GB

Qwen3.6-35B-A3B on the GPU (30B-class MoE)

ESTIMATE RAM or VRAM use at Q4 with a small KV cache, from the litecode catalog. Any GGUF model works. Hardware guide.

How it works

Five steps, all on your machine

Detect

It reads CPU, RAM, disk type, GPU and VRAM.

Recommend

It picks models that fit, with room for your other programs.

Ask

It shows the download size and downloads only after you agree.

Tune

It measures threads and decoding modes and keeps the fastest.

Work

It reads, edits and tests your code. You approve every change.

Features

A full coding agent that stays on track

Small models make more mistakes, and big models cost time. litecode adds structure to both: a strict tool protocol, grammar-forced output, and tests after every change.

01Agent with tools

Read, search, exact edits with a diff preview, run commands, git. Modes: ask, auto, plan, readonly. /undo any edit.

02Memory that grows

Every session is saved to disk. Your preferences apply to all projects. Rated answers come back as examples. Details.

03Skills

Save a workflow as a SKILL.md. Run it with /name. Let the model draft one with /skill make.

04MCP client

Connect Model Context Protocol servers. Reads the common .mcp.json config format.

05Speed on any machine

Prompt cache saved per project, n-gram speculative decoding, GPU offload, and litecode tune to pick by measurement.

06Plain answers

Answers follow ASD-STE100 Simplified Technical English: answer first, short sentences, exact values, no filler.

07Plan mode

/plan inspects without editing and gives numbered steps. You approve, then it runs them.

08Benchmark yourself

litecode eval runs 30 real tasks in 3 levels and scores each model on your machine.

09Editor autocomplete

litecode fim serves fill-in-the-middle completion. A small VS Code extension is included.

Quick answers

Common questions

?
Can I run an AI coding assistant without a GPU?Yes. litecode runs on CPU only. A 2019 Ryzen 3 with 16 GB RAM generates about 15 tokens per second with the default 2B model.
?
Does it use my GPU if I have one?Yes. Setup installs the CUDA or Vulkan engine and moves the model to the GPU. With 24 GB of VRAM it runs 30B-class models.
?
Is my code sent anywhere?No. After setup litecode makes no network calls. Code, prompts and memory stay on your disk as plain files.
?
Is it only for coding?Coding and security come first. It also writes emails, summaries and docs, answers study questions and handles CSV, JSON and logs.
?
What does it cost?Nothing. litecode is MIT licensed and the models are open-weight. More answers.

Install

One command

Needs Python 3.10 or newer. On Windows the installer offers to install it. litecode gets its own environment and the litecode command. The setup wizard then downloads the engine and the models that fit your machine, after it asks.

Windows · PowerShell

irm https://raw.githubusercontent.com/konkomaji/litecode/main/install.ps1 | iex

Linux · macOS

curl -fsSL https://raw.githubusercontent.com/konkomaji/litecode/main/install.sh | sh

Then run litecode tune once, and litecode in any project folder. Read the Quickstart or the full User guide.