CyberFlex / CodingDude

CodingDude 0.4.0 · open source · by Cyberflex

An AI coding agent that has to prove it reviewed its own diff.

CodingDude plans, designs, builds, reviews and ships work in your repository, on your own machine, with the AI model you choose. It asks before anything risky and shows proof before it calls anything done.

  • Self-hosted, single-user
  • Python 3.12+
  • Any OpenAI-compatible model
  • AGPL-3.0-or-later
A risky step stops and waits for your approval.

Why CodingDude

Your agent, your models, your machine.

Hosted coding assistants keep your code and keys in someone else's cloud, tie you to their models and act on your repository with little say from you. CodingDude is built the other way round.

Private by design

Everything runs on your machine. The only thing that leaves it is the calls to the AI model you chose.

Any model you like

DeepSeek, OpenRouter, OpenAI, Anthropic or a local model through Ollama, chosen per job, with automatic fallback.

You stay in charge

Safe steps run on their own. Anything risky waits for your approval.

Proof, not claims

Nothing counts as done without test, lint and build results to show for it.

The review gate

Review is a gate, not a habit.

CodingDude can't publish code until it has reviewed exactly what it is about to push. If anything changes after the review, even one line, it has to review again.

  1. Check the standards

    Does the code follow the way your team writes software?

  2. Check the spec

    Does it do what the ticket asked for?

  3. Fix and re-review

    Real findings are fixed, and the changed code is reviewed again.

  4. Publish

    Only reviewed work is pushed, and only with your approval.

A recorded pre-push review: Standards found three issues and fixed two, Spec found none, and the push is allowed only while the diff matches.
Every push comes with a record of what was reviewed and what was found.

Before it builds

Understand, design, then build.

Good engineers don't start typing straight away. Neither does CodingDude.

Graph panel showing the call graph for a file, with dead-code warnings and a control-flow view.

Map your code

An architecture map, call graphs, which tests cover each file, and dead code. Ask "where is authentication enforced?" and get answers linked to the exact lines.

A lo-fi wireframe of a Providers screen, with annotated callouts for each region.

Sketch the screens

Ask for a screen and get a wireframe, with the flow between screens. Edit it, then press implement to build exactly what you approved.

Trace panel listing the run's events, with file checkpoints and a rewind button for each.

Undo anything

Every edit is snapshotted first, with one-click rewind. Suggest mode turns edits into proposals you accept or revert.

Spend control

A budget it can't overspend.

Set a limit per run or per session. CodingDude checks it before every call to the model, so the work stops at your limit instead of you finding a surprise bill afterwards.

  • Covers everything the agent does, including the helpers it hands work to.
  • Long sessions keep going. The conversation is condensed automatically instead of running into the model's limit.
Cost panel showing requests, tokens and cost, broken down by model and by function.
Every request, token and cent, broken down by model and by job.
Planning panel with PRDs, tech designs and roadmap items, each with its status.
The roadmap and tickets you and the agent work from.

Planning

Work that outlives the chat.

Requirements, designs, a roadmap and tickets are kept as real records in your project, not lost in a chat history. You and the agent work from the same plan.

  • Proof, not promises. A ticket step is only marked done with evidence that it works.
  • Pick up where you left off. Sessions, notes and decisions carry over from one day to the next.

Approvals

Safe actions run. Risky ones wait for you.

Reading code and planning happen on their own. Pushing code, installing packages and deleting files always stop and ask, on one clear approval card next to the change.

  • Approve repeat commands once, such as your test runs, for the rest of the session.
  • Your keys stay out of reach. Commands the agent runs are started without your API keys and passwords.
An approval card asking to run a command, with Approve once, Approve all matching and Deny, above the diff it relates to.
The change, and the one decision waiting for you.

In the box

A workspace you arrange yourself.

Dock, split or float any panel, or pop one out onto a second monitor, and save layouts for reviewing or debugging. Follow each run live: what the agent is doing, what changed, where the time went and what it cost.

The built-in editor comparing a changed file with the last commit.
The built-in editor, showing a change against the last commit.
Waterfall panel listing recent runs and a timeline of each tool call.
Where the time went, step by step.

Fix until green

Runs your tests after a task and feeds any failures back to the agent, for up to three rounds.

Built-in security scan

Bandit, Semgrep and secret detection check the work, and findings stop it going further.

Your house rules

Standing rules, project notes, custom skills and review checklists, all edited inside the app.

Memory

Remembers facts, solved problems and decisions from one session to the next.

Team of helpers

Hands research and reviews to helper agents working in parallel, without touching your code.

Scripts and CI

Runs without the UI on a fixed budget. Record a real session and replay it later as a regression test.

Get started

Up and running in three steps.

CodingDude runs on your own machine, against your own code. There is nothing to sign up for. On Windows, the setup script installs everything else for you.

  1. Install Git and Node.js

    Both are free: Git and Node.js 20 or later.

  2. Download CodingDude and run the setup

    command prompt
    git clone https://github.com/cybersenses/codingdude.git
    cd codingdude
    setup-codingdude.bat
  3. Start it

    Double-click start-codingdude.bat. CodingDude opens in your browser; add your AI model key and you're ready.

What the setup installs

  • Python 3.12 and CodingDude itself, through uv
  • Semgrep, for the built-in security scan
  • The GitHub CLI
  • The code-map and GitHub connectors
  • The web app, built on your machine

Safe to run again at any time. You'll also need an API key from DeepSeek or any OpenAI-compatible provider.

On Linux or macOS? Follow the getting started guide →

Licence

Free and open source. Your code stays yours.

CodingDude is free software under the AGPL-3.0-or-later licence, and everything it writes for you belongs to you. Read the licence terms.

Commercial licence

Building it into your own product?

If the open-source licence doesn't suit your business, a commercial licence is available from Cyberflex.

Talk to Cyberflex

Community

Questions, bugs and ideas.

The CodingDude mascot giving a thumbs up beside the steps plan, design, implement, review and deploy.

CodingDude is built by Cyberflex in Mogale City, South Africa. You bring the AI model; CodingDude brings the discipline.