Numex Codex

A coding agent that thinks in Turkish — for your editor, terminal and browser.

Ask. It writes, runs and verifies — until the job is actually done.

🇹🇷 Türkçe · 🇬🇧 English

---

Numex Codex is the coding agent of Numex AI, Türkiye's Turkish-first AI. Tell it what you want in plain language; it plans, writes, runs and tests the code, and only says "done" once it can prove it works. For complex work it doesn't act alone — it convenes a Council of four specialists.

📌 This is Codex's showcase, documentation and community repo. Downloads: codex.numexai.com.tr. Open-source SDK: numex-sdk.

🚪 Four doors, one agent

	Where	Install
🧩 Codex IDE	Windows desktop (VS Code based)	Download — v2.9.59, ~162 MB
🧩 VS Code extension	Your existing VS Code	Install the extension + sign in via CLI
⌨️ Codex CLI	Terminal	npm i -g @numexai/cli (v3.3.10)
🌐 Codex Web	Browser	Nothing to install

🏛️ The Council — four specialists, one task

Most coding agents grade their own homework: one model plans, writes, tests and declares success. Codex hands complex tasks to a Council where every member has one job and limited permissions:

Member	Job	Can write?
📐 Architect	Folder structure, data model, per-file function plan	❌ Plan only
💻 Coder	Writes complete, working code — no stubs, no TODOs	✅ The only writer
🛡️ Auditor	First asks "did we build the right thing?", then tests and verifies	❌ Read & report only
🎨 Designer	Polishes UI/UX	⚠️ CSS/HTML only

• Single-writer rule — only the Coder writes code, so agents never collide.
• Off-topic plan check — if the Architect's plan drifts from your request, the Coder follows your request, not the plan.
• Right-sized team — a one-line fix gets no council; a full-stack system gets all four. Token budget is checked up front (trivial ≈ 500 tokens … enterprise ≈ 30,000).
• Summon it yourself — "konseyi topla" ("gather the council").
• Honest report — a plain-language summary of who did what, including anything left unresolved.

→ Full Council write-up (Turkish)

🚦 No "done" without proof — FinishGate

A task only closes with live evidence: an HTTP 200 from the running server, passing tests, or the expected DOM in a headless browser. Otherwise the agent returns to a self-healing loop. If it gets stuck, a stuck-detector hands control back to you.

🔀 Four modes

Chat (explain, answer) · Plan (no edits, just a roadmap) · Autonomous (plan → approve → apply → verify) · Bug Hunter (root cause from logs and output).

⌨️ CLI highlights

numex "build a dark-themed todo app"
numex plan "add a payment module"      # plan only
numex mission "make the tests pass"    # long-running autonomous mission in sandboxed worktrees
numex undo                             # roll back the last turn
numex --offline "…"                    # local Ollama, no internet
numex uzak install                     # control your PC from your phone

Surgical diff-based patching · background process error interception · BM25 + AST code index · MCP client · live tree progress.

🧠 Transparent memory

Every project gets a readable .numex/ folder: RAG index, audit trail of every action, undo checkpoints, learned errors, mission state and token usage. → docs/numex-klasoru.md

🤝 Community

Bugs and ideas → Issues · Show what you built → Discussions · Security → destek@numexai.com.tr