AbdoCode: A Local-First Coding Agent Where the Model Proposes and the Kernel Decides
AI coding agents are becoming powerful enough to read files, edit code, run commands, and interact with development environments. But as agents gain more capabilities, one question becomes increasingly important:
How much control should an AI agent have over your machine and your codebase?
This question led us to build AbdoCode, a coding agent for Windows that runs on your machine and is designed around a simple principle:
The model proposes. The kernel decides.
What is AbdoCode?
AbdoCode is a local-first coding agent built by TechnologyKSA, the Saudi technology company behind Mubarmij.
It works with different types of models:
Local models through Ollama.
Cloud models using your own provider API key.
The agent can read, write, edit, and run code inside a project. However, actions do not go directly from the model to the operating system. They pass through a controlled execution layer designed to make changes more visible, reviewable, and reversible.
Why a control kernel?
A language model can suggest a useful change. It can also misunderstand a requirement, repeat a broken edit, or generate a command that has a larger effect than intended.
AbdoCode uses a small Rust-based kernel as a control point between the model and the machine. The goal is not to remove the model from the workflow. The goal is to make its actions easier to inspect and safer to manage.
The workflow is built around several ideas:
Preview changes before trusting them.
Review the result instead of relying only on the model's explanation.
Keep a rollback point for each turn.
Apply stricter safeguards when the model behaves unreliably.
Keep provider keys and secrets on the local machine.
Features for everyday development
AbdoCode is designed to support more than code generation. It can work with project files, screenshots, browser workflows, and external tools.
The current project includes capabilities such as:
Reading, writing, editing, and running code.
Previewing changes before applying them.
Rollback points for agent turns.
Review lenses for correctness, safety, and tests.
Browser tools for reading pages, finding elements, filling forms, and handling overlays.
An Edge-based agent browser and a browser extension for working with your own tabs.
Local model support through Ollama.
Cloud model support with user-provided keys.
Windows Vault storage for secrets.
Optional routines and local-network remote control.
Connectors for selected external tools through OAuth and MCP.
A small ecosystem around the agent
The main application is accompanied by additional repositories:
Project templates
The AbdoCode templates repository contains 100 versioned project starters. The templates cover combinations of frameworks, databases, ORMs, styling systems, and common application features.
They are intended to help developers start with a structured project instead of an empty directory.
Add-ons
The AbdoCode add-ons repository contains optional instruction skills and a read-only Gmail MCP connector.
Add-ons are installed explicitly. Packages can be reviewed, verified using their published hashes, installed, and enabled by the user. The Gmail connector is read-only in the current version and cannot send, delete, or modify messages.
Local-first does not mean local-only
A local-first workflow should not force every developer into one model or one provider.
Some developers prefer local models for privacy, offline workflows, or experimentation. Others need the capabilities of cloud models. AbdoCode is designed to support both approaches while keeping the execution and credential boundaries on the user's machine.
This separation is important: the model can be replaceable, while the control layer remains part of the development workflow.
Source availability and licensing
The AbdoCode source and documentation are published on GitHub for reading and local builds. Before using the project commercially, review the published license and commercial terms. The repository states that commercial use requires a written license.
This distinction matters. A project can publish its source for inspection and local building without granting unrestricted commercial rights under every definition of open source.
Try AbdoCode
The project is available here:
AbdoCode on GitHub
AbdoCode templates
AbdoCode add-ons
We would love to hear from developers building with AI agents:
What makes you trust an AI coding agent?
Do you prefer local models, cloud models, or a combination of both?
Which safeguards are essential before an agent can edit and run code?
What would you want an AI coding agent to show you before it changes your project?
code-ksa
/
abdocode
عبدو كود — وكيل برمجة يعمل على جهازك بنواة Rust، مع أيّ نموذج محلّيّ أو سحابيّ. AbdoCode: a local coding agent for Windows by TechnologyKSA, maker of Mubarmij.
AbdoCode — عبدو كود
A coding agent that runs on your own Windows machine, with a small deterministic Rust kernel, and works with any model: local through Ollama, or cloud with your own key.
⬇ Download the latest Windows installer · For AI agents · Drive it from your own agent
Why we built it
AbdoCode is a product of TechnologyKSA (technologyksa.com), the Saudi technology company behind Mubarmij (moparmeg-ksa.com). We wanted an agent that lives on the machine, keeps the keys in the machine, and treats the model as an advisor rather than an owner.
The idea is simple: the model proposes, the kernel decides. Every effect on your machine goes through one gate, is checked against the mode you chose, and is judged by its receipt, not by what the model says about it. Small models get stricter rails, strong models get thinner ones, automatically.
Tags
Top comments (0)