Punani AIpowerd · naga codex
Punani
🇩🇪 Made in Germany · HamburgFree · Open Source · Production-grade permissions

The coding agent that asks before it touches your repo.

A Cursor-level AI coding agent in the CLI and the web browser, with $0 subscription lock-in. Your keys stay on your machine. Free models first.

↗ Open the app↓ Install in 60s✉ Join the list
punani · latest build
This build
  • Read
  • Edit
  • Diff
  • rm -rf ./build

Open the app takes you to sign-in steps — the agent runs on this computer at http://127.0.0.1:4096 after punani web. New here? Use Install first.

8model providers
34automated tests
0surprise writes
MITopen source

Why Punani

Top-tier AI. No monthly lock-in.

Built for developers who want to code with top-tier AI for free instead of paying monthly $20/mo subscriptions.

◎

Built for the free-model crew

For builders who love free tiers, run local Ollama, or simply cannot justify $20/month agents. OpenRouter free models, Groq and Cerebras speed, NVIDIA NIM credits, FreeLLMAPI failover — ranked so a 429 sends you to the next option, not a paywall.

Start with punani --init, pick a free model, ship real code with tools, MCP, and previews — not a demo chatbot.

⬡

Bring OpenAI, Grok, or any key

Already paying for OpenAI, Anthropic, Groq, or routes through OpenRouter? Paste your key once — stored on your machine or only in the browser if you prefer. Swap models from live catalogs with pricing shown up front. Punani does not resell tokens; it orchestrates your stack.

One agent loop in terminal and punani web. Same rules. Same diffs. Your choice of brain.

Built like a paid product

Looks expensive. Costs nothing but your API key.

A full agent loop — read, search, edit, run, generate image and audio — with permission cards, live preview, and resumable sessions. No black-box rewrites.

punani · permission to touch
Punani AI in the browser — a permission card blocking a destructive shell command before it runs

What you get

Four surfaces. One disciplined engine.

Same permission rules in the terminal and the browser. Last match wins. Free providers surface first when something rate-limits.

◆

Terminal coding agent

A full agent loop in your terminal: slash commands, resumable JSONL sessions, diff preview before every write, and error messages that name the fix.

◆

Browser command center

The same engine in the browser over SSE, with a native folder picker, a live preview pane and API key storage on disk or in the browser only.

◆

Model freedom

Eight model providers behind one agent: OpenRouter, NVIDIA NIM, Groq, Cerebras, OpenAI, FreeLLMAPI, Ollama and LM Studio, with live catalogs and real pricing.

◆

Permission-first editing

Every edit shows a diff before it lands, and anything that can lose work — a file that would be overwritten, a command that deletes — stops at a permission card and waits for you. Nothing is rewritten behind your back.

Eight model providers, one agent. Free options are ranked first and rate limits fail over instead of stopping.
ProviderKindWhat it gives you
OpenRouterFree tier firstHundreds of models, free variants ranked ahead of paid.
NVIDIA NIMFree creditsHosted open models on NVIDIA's free tier.
GroqFast free tierVery fast inference on the free allowance.
CerebrasFree tierFast inference for open-weight models.
OpenAIBring your keyYour own key, your own bill, your choice of model.
FreeLLMAPIFailoverAutomatic fallback when a provider rate-limits.
OllamaLocalRun open models on your own hardware, fully offline.
LM StudioLocalLocal server with live model discovery.

Install

Ship in one terminal.

Global CLI, wizard for provider + key, then pick a free model and go.

$ npm i -g punani-powerd
  • installnpm i -g punani-powerd
  • setuppunani --init
  • terminalpunani
  • browserpunani web — open it now

    • Agent: Punani-Power-Agent· Landing: Punani-AI· punani.agency

    Free tier, honest defaults

    OpenRouter free models, NVIDIA NIM credits, FreeLLMAPI failover — ranked so you are not stuck behind a 429.

    $ npm i -g punani-powerd$ punani --init provider openrouter model openrouter/minimax/minimax-m3.1-flash:free ✓ ready — ctx 1.0M · tools ✔ · free

    Questions

    What people ask before they install.

    The short answers. Longer ones live in the about page and the repository.

    Is Punani AI really free?

    Yes. There is no subscription to us and no token resale. Punani ranks free tiers first — OpenRouter free models, NVIDIA NIM credits, FreeLLMAPI failover — so a rate limit sends you to the next option instead of a paywall. If you already pay for OpenAI, Anthropic or Groq, paste your own key and it runs on your model instead.

    What does "asks before it touches your repo" mean?

    Every edit shows a diff before it is written, and anything that can lose work stops at a permission card and waits for you. A destructive command such as rm -rf ./build stops at the card. The ordinary work — reading, writing, editing, running a command — does not interrupt you. Nothing is applied in the background.

    Which models can Punani AI use?

    Eight providers: OpenRouter, NVIDIA NIM, Groq, Cerebras, OpenAI, FreeLLMAPI, Ollama and LM Studio. Punani reads live catalogs, shows real pricing, and lets you swap models without changing the agent.

    Can I run it fully offline or on my own hardware?

    Yes. Point Punani at a local Ollama or LM Studio instance and the agent runs on your machine with no cloud calls. Your API keys are stored on your machine, or in the browser only if you prefer that.

    Does it support MCP and the Model Context Protocol?

    Yes. MCP servers can be registered with the agent, so tools, resources and prompts from other projects are available inside the same permission-first loop.

    Is Punani AI open source?

    Yes, MIT licensed. The landing page and the agent engine are both public on GitHub: Punani-AI for the site and Punani-Power-Agent for the agent itself.

    Who builds Punani AI?

    Maurice Holda, working through Naga Codex in Hamburg, Germany. The same studio runs Naga Apparel, an urban streetwear label from Hamburg shipping worldwide. Both sit under the Naga brand family, and Punani AI is the software product.

    Which platforms does it work on?

    Terminal and browser. The terminal gives you slash commands and resumable sessions. The browser gives you a chat loop, a folder picker and a live preview pane. Same permission rules, same diffs, in both.

    Contact

    Talk to the person who builds it.

    Punani AI is made by Maurice Holda in Hamburg, Germany, through Naga Codex. The same studio runs Naga Apparel, an urban streetwear label from Hamburg shipping worldwide. Questions about the agent, the agency or the label go to the same inbox.

    chosenfewrecords@hotmail.de · Hamburg, Germany · About the studio