A self-hosted web app that onboards engineers to a codebase — or explains a large PR — by generating a tutorial-like, self-paced interactive experience: "a 1:1 session with an expert in the codebase."

Paste a GitHub repo (or PR) URL and Code Exploder analyzes it with locally hosted LLMs, then serves:

• an intro to what the code is and what the product does;
• a whiteboard architecture tour — diagrams that progressively reveal complexity, walking through data flows, user interactions, and product scenarios;
• a description of how the code is built, tested, and released;
• optional quizzes per section to check understanding;
• an interactive Q&A chat with the virtual expert, grounded in the analysis knowledge base, with file/line citations.

Everything runs on your own hardware: .NET 10 + React, PostgreSQL (+pgvector), and a local Ollama endpoint. Inference is queued so a single shared GPU is never overwhelmed.

The core product (M0–M4) works end-to-end today. Full descriptions in docs/08.

The LLM defaults to qwen3-coder:oc via an OpenAI-compatible endpoint (Llm__BaseUrl, default http://localhost:11434/v1); the embedding/generation GPU-co-residency budget is documented in docs/06.

One command installs the toolchain, a container runtime for Postgres, and a local Ollama serving the models:

scripts/setup-macos.sh

It's idempotent (safe to re-run) and installs, via Homebrew: .NET 10 SDK, Node, the GitHub CLI, Docker (Colima if you don't already have a daemon), and Ollama — then pulls the generation model qwen3-coder:30b and the embedder nomic-embed-text, and starts Ollama with a large context window.

Model note. Production runs qwen3-coder:oc, a Modelfile that only widens the context window of the official qwen3-coder:30b for a specific GPU. Locally we use the base qwen3-coder:30b and widen its context through Ollama's OLLAMA_CONTEXT_LENGTH setting instead (the setup script sets this). The 30B model is comfortable on Apple Silicon with ≥32 GB unified memory; on a smaller Mac re-run with e.g. OLLAMA_CONTEXT_LENGTH=32768 scripts/setup-macos.sh (large-repo synthesis quality drops as the window shrinks).

The setup writes .env.local (gitignored) with LLM_MODEL=qwen3-coder:30b, which dev.sh picks up — so plain ./dev.sh uses your local model with no extra flags.

Start the complete local development stack with:

./dev.sh

The script starts PostgreSQL, the gateway, both workers, the orchestrator, and the Vite development server, then opens the app at http://localhost:5173. Press Ctrl+C to stop everything it started.

Analyzing a private repo. The app clones repos by URL with credential prompts disabled, so private repos need your credentials. Point dev.sh at one you can access and it wires your gh login into the workers' git (process-scoped — your global ~/.gitconfig is untouched), then prints the URL to paste:

gh auth login                       # once, if you haven't
REPO=your-org/your-private-repo ./dev.sh
# → Paste this URL in the app: https://github.com/your-org/your-private-repo

REPO accepts owner/name or a full github.com URL, and can live in .env.local instead. Both repo and PR analysis of private repos then work with your token.

To run the services individually:

docker compose -f deploy/compose.yaml up postgres -d   # Postgres 17 + pgvector on :5433
dotnet run --project src/CodeExploder.Gateway           # API + hub on :5080 (DevBypass auth)
dotnet run --project src/CodeExploder.Workers.Analysis  # deterministic-analysis worker
dotnet run --project src/CodeExploder.Workers.Llm       # generation + embedding lanes (needs Ollama)
cd webui && npm install && npm run dev                  # Vite dev server, proxies to :5080

Tests: dotnet test codeexploder.slnx (queue join semantics run against a throwaway Testcontainers Postgres — Docker required) and cd webui && npm test.

The initial system design study lives in docs/:

MIT