Hello. I'm back again! I apologize for posting so frequently. I'll hold back a bit after this post! 🙇♂️🙇♂️🙇♂️
"How to Use GPT + Local LLM Together - 3 Role Pattern, Code Included "This is about properly implementing the concept and code from that post in plugin form! (I made it on Mac, and I'm not sure why it works on Windows/Linux (haven't tested), but thank you anyway. It's probably just Python scripts, right? hehe)
Among the AI Damoang community members using Claude Code or Codex CLI on a 20-dollar plan, this is a tool I made for those who are concerned about "quotas running out quickly" or "I want to run a local LLM together, but how do I connect it?" to try out. (I have no other intentions! I use different tools primarily hehe)
Link: https://github.com/hang-in/tunaLlama
What is a Plugin First?
Both Claude Code and Codex CLI can plug in external tools as "plugins". Once installed, the agent will automatically call it whenever needed during tasks. You can think of it as a package where the agent automatically uses the tools (MCP servers), work instructions (Skills), automatic triggers (Hooks), and shortcut commands (Commands) inside it. If MCP is a tool connection standard, a plugin is a unit that bundles tools with usage instructions and deploys them all at once. Users don't have to keep saying "use this tool" every time. tunaLlama is also a type of that plugin.
What Does It Do?
It's divided into three roles.
Role | Who | What They Do |
|---|
Architect | Claude / Codex (Subscription) | Break down tasks, validate, integrate |
Developer | Local LLM (Ollama / LM Studio / Ollama Cloud) | Write actual code, self-review |
Reviewer | Architect (Same session) | Final verdict |
In simple terms, it's structured as "let the local LLM do the heavy coding, and use the subscription model only for the thinking parts (breaking down and validating)". The subscription quota runs out slower, and at the same time, the subscription model supplements the context that's insufficient for a local LLM alone, so the result quality can also improve.
Quantitatively Measured Part
When running mid-size local LLM alone vs. after Architect organizes and passes context, the comparison shows +0.58 ~ +0.64 (3 model validation). This means that with the same local LLM, if context is well organized and passed, the results improve meaningfully. However, this measurement was done on "a test set I prepared in advance with task scenarios that often occur in reality", so it's a separate question whether the same results will appear in your actual workflow.
How Much Quota Savings?
To be honest about this part - since Anthropic / OpenAI's quota calculation formula is confidential, I can't give you numbers like "X% savings". "Better than just using it as-is", and I recommend confirming through actual use. Actually, the code since v0.2.0 was developed through dogfooding, and the usage didn't noticeably increase between v0.2.0 and v0.5.0 (it might be because it's Python. I'll test this more on other projects!)
Korean Language Support
I've attached Kiwi morphological analyzer and several other things. I reused what was already made in secall. hehe When searching for past work, Korean queries are also indexed at the morpheme level. However, neologisms and technical terms that Kiwi cannot process can affect search quality.
5-Minute Installation (Summary)
Turn on the terminal agent and drop a single line in the session and you're done:
Install following the INSTALL.md from https://github.com/hang-in/tunaLlama
For those using ollama cloud, just prepare the API key in advance (I recommend it - it's about 30,000 won worth of happiness) After that, the agent will automatically install dependencies, configure .env, register the plugin, and verify it step by step. For those using gemma4, qwen3.6, or similar with ollama or lmstudio locally, the terminal agent will just handle it automatically.
For Claude Code, once installed, you can start fresh or use /plugin-reload,
For Codex, you'll need to install it once with /plugin 😁
Or if you prefer completely manual installation, refer to the README.
Limitations
It's production v0.5.0, but real-world usage data is still being collected.
Quota savings is based on subjective data (mentioned above)
Won't work without a local LLM environment (Ollama, etc.) (27b ~ over 32b recommended, if you use Ollama Cloud, Kimi or GLM would be even better)
13 MCP tool descriptions add approximately 1,633 tokens to each conversation system prompt. This cost is quality-related, so it's something to accept (but it's much cheaper(?) than doing everything with Claude Code/Codex)
Who Would Benefit from Using This?
Claude Code / Codex 20-dollar subscription users (quota management motivation)
Those who already have Ollama local / Ollama Cloud / LM Studio environment
Those who handle Korean language work (Kiwi integration)
Conclusion
It's not a spectacular tool. I made it for my own use, and I'm just saying let's use it together if there are some community members in similar situations. I have no intention of promoting it elsewhere. This is the only thing, however modest, I can contribute to the Damoang community. 😁 Since it's MIT licensed, you can fork it and modify it to fit your workflow. Feedback, issues, PRs, and stars (⭐️) are all welcome. If you want feature additions, modifications, or additional MCP integrations, feel free to let me know via comments or issues! 🥕🥕🥕
