Mix local LLMs, Claude Code, Codex, Gemini and more in one SDLC pipeline (open source) [P]

Problem: one model for everything

A lot of AI coding tools assume one model should do everything: understand the task, write the code, review it, and decide whether it is correct. I ended up building something around the opposite idea. Instead of one model doing the whole software development lifecycle, every stage can use a completely different model. That means you can mix local LLMs and hosted models however you want.

Example model combinations

• DeepSeek-R1 running locally for planning
• Claude Code for implementation
• Qwen-Coder running on Ollama as a reviewer
• Gemini CLI for another project
• Codex for only the coding stage
• or any combination you prefer

Each stage is independent, so you're never locked into one provider or one model family. The coding stages don't even require API keys if you already use tools like Claude Code, Codex, Cursor, Aider or Gemini CLI. AutoDev Studio simply drives them headlessly through their existing login. Meanwhile other stages can run entirely on local models through Ollama or any OpenAI-compatible endpoint.

Design goal: different reviewer than author

The reviewer is intentionally a different model family than the author, so the same model never gets to approve its own code. That was one of the design goals from the beginning.

Pipeline stages

The pipeline itself looks like this:

• PM agent runs a clarification loop and produces implementation tickets
• optional Jira sync
• Dev agent implements on an isolated branch
• QA runs the repository's real tests
• reviewer validates the diff
• bounded revise loop if QA or review fails
• opens a real pull request for a human to merge

Every stage records its own tokens, runtime and cost.

Motivation and example usage

I originally built this because I wanted to experiment with using specialised models where they actually make sense instead of asking one model to do everything. Some examples I've been using:

• local reasoning model → planning
• Claude Code → implementation
• local Qwen-Coder → review
• local embedding model → repository indexing

Benchmark results

The benchmark that ended up surprising me was that, on two large Python repositories (35k and 82k LOC), the tuned pipeline beat a cold Claude Code run on all six well-localised tasks, between 7% and 75% cheaper by avoiding repeated repository exploration. It definitely doesn't always win. Tiny one-line edits are often cheaper with a single cold agent, and I included those failures in the benchmark as well.

Key features

• works with both local and hosted models
• supports OpenAI-compatible endpoints
• runs on free local models if you want a completely offline workflow
• language-agnostic pipeline
• live dashboard with streamed agent logs
• per-stage cost and runtime tracking
• MIT licensed

Repo: https://github.com/krishagarwal314/autodev-studio

Call for feedback

This is still an early side project, so I'd genuinely love feedback-especially from people experimenting with local models. If you've got ideas for better model combinations, want to try different local LLMs for individual stages, or find places where the pipeline falls over, I'd really appreciate an issue or discussion. And if you think the project is interesting, a star would mean a lot.