DeepSeek + REA Cloned Claude Code for FREE (Full Build)

DeepSeek + REA Cloned Claude Code for FREE (Full Build)

Title history
  • DeepSeek + REA Cloned Claude Code for FREE (Full Build) — observed on 20261115-1739 UTC (YouTube)
🎙 Outskill 👥 29K 📅 November 11, 2026 ⏱ 26 min 👁 284K 📄 tutorial 🧭 2026-11-15
Available in: English (current) Français

Keywords

free AI modelsreverse engineeringcoding agentDeepSeekREA

Summary

The video demonstrates rebuilding Claude Code from scratch using entirely free AI models: DeepSeek V4.1 Flash and GLM 5.3 Flash, with the help of REA (Reverse Engineer Anything). The presenter walks through installing OpenCode and DeepSeek Harness, configuring REA to study the Claude app and its engine, and then using AI agents to write a new coding tool named Kiln. The process includes a warm-up by cloning an award-winning 3D website, then reverse-engineering Claude Code’s features (permissions, MCP, hooks, helper agents) and rebuilding them. The build uses a self-check automation script that drives the Kiln window to verify each feature, achieving 77/77 initial tests, then increasing to 102/102 after fixing missing features. The video highlights common pitfalls like free model rate limits, the need to verify AI claims, and legal/ethical constraints. A head-to-head comparison between Kiln on free models and the real Claude app is shown, with the presenter leaving the verdict to the viewer. Practical lessons are shared about writing rules in AGENTS.md, saving versions with Git, and treating ‘done’ as a claim that requires proof.

177 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides significant practical value for users who want to use AI coding agents without paid subscriptions. The step-by-step approach is exhaustive, showing every error and fix, which strengthens the argument that the process is reproducible. The justification for using REA and free models is clear: to demonstrate that a capable coding agent can be built without paying for proprietary tools. The argumentation is solid, as the presenter supports claims with concrete evidence: the self-check script passes, screenshots of rebuilt websites, and side-by-side comparison. The video also addresses realistic limitations (free tier rates, privacy), which adds credibility. However, the argumentation is primarily anecdotal and not a systematic comparative study; the side-by-side test is subjective, and the video does not provide metrics like token usage or completion time precisely.

Scientific Rigor, Source Quality, Title Accuracy

The video uses real tools and provides links to official sources: REA GitHub, OpenCode, DeepSeek Harness, OpenRouter, NVIDIA, and the resource library. These are mentioned in the description and are relevant to the tutorial. The scientific rigor is moderate: the presenter explicitly states that they are not copying code and that the build is private learning, which respects Anthropic’s policy. However, the reverse engineering approach raises legal questions that the video acknowledges. The title is accurate as it clearly describes the content. The video does not present original research but a practical demonstration, yet it is well-documented with explicit prompts and troubleshooting. No external sources are cited beyond the tool links, and no contradictory evidence is presented, which limits the depth of discussion.

267 words

Title / Content Match

The title exactly matches the content: the video shows how to clone Claude Code using free AI models (DeepSeek) and REA, providing a full build walkthrough.

Quality & Reliability

7/10

The video is transparent, showing failures and corrections, using real tools and verifiable outputs. It is a practical tutorial rather than a scientific study, but it provides a reproducible workflow. Claims about free models running out and privacy are honest. The demonstration of building Kiln with 102 passing self-checks is impressive but lacks external verification.

Key Moments

Cited Sources

  • REA GitHub — Official repository of REA, the reverse engineering tool used throughout the video.
  • OpenCode — Free AI coding agent application used as one of the bodies for the AI.
  • DeepSeek Harness — Free AI harness app used for the website reverse engineering part.
  • OpenRouter — Provider of free AI model APIs, used to get free keys. Also mentioned for model selection.
  • NVIDIA free API keys — Provides free API keys for models like GLM 5.3 Flash, used as alternative provider.
  • Resource Library — Place where the presenter says prompts and setup docs are available.
  • Outskill Mastermind Promo — Promotional link for the 8-hour live AI mastermind mentioned in the video.
  • Next video on REA — Related video 'How to Use REA GitHub to Reverse Engineer Anything'.

Contribution & Novelties

The video’s original contribution is a full, reproducible workflow for building a coding agent (Kiln) using only free AI models and the REA reverse-engineering tool. It demonstrates that Claude Code’s core features (permissions, MCP, hooks, helper agents) can be recreated without paid APIs, and it surfaces practical lessons about self-checking, recovery from model rate limits, and the importance of verifying AI output. The step-by-step prompts, error handling, and the use of the robot hand self-check are novel in their clarity.

Pour aller plus loin :

  • REA on GitHub — The tool at the heart of the video: a collection of 229 reverse-engineering tools. This is the primary reference for understanding how to study and clone applications.
  • Ghidra official site — The NSA’s free decompiler used in the video to analyze native executables. A key component for reverse engineering desktop apps.
  • OpenCode documentation — The official documentation of the free AI coding agent used as the body. Useful to understand configuration and model selection.
  • three.js official site — The JavaScript 3D library used to rebuild the website in the warm-up. Relevant for understanding the technical stack mentioned.
  • DeepSeek official site — The models provider used in the video. More information about DeepSeek V4.1 Flash and free tier limitations.

207 words

Radar Profile

The scores show a high quantity of information (9) and a good technical level (8), indicating a dense and advanced tutorial. Quality and reliability are slightly lower (7 and 7), reflecting the lack of external verification and the practical nature. The overall profile suggests a valuable, hands-on resource for experienced users, but not a rigorous scientific reference.

Reliability 7/10

💬 Très positif — sur les 30 commentaires analysés, la majorité exprimait de l'enthousiasme et de la reconnaissance pour la démonstration, avec quelques demandes de ressources supplémentaires et des remarques techniques, sans critique majeure.