Recipe: Repo-aware coder
Build a local coding agent that indexes your entire repository, then answers questions and writes patches with full context — no cloud dependency, no token limits on codebase size.
Ingredients
- Ollama running codellama or deepseek-coder
- Tree-sitter grammars for your target languages
- FAISS or Chroma for vector search
- Nimbus SDK for hardware-bound licensing
Steps
- 1
Index the repo
Walk every source file, chunk by function/class using tree-sitter AST queries. Embed each chunk with your local model and store in FAISS.
- 2
Build the retriever
On query, embed the user prompt, run ANN search, and collect the top-K chunks. Concatenate into a single context window.
- 3
Prompt and patch
Send context + user request to the LLM. Parse the response for unified diffs. Apply patches with a dry-run first.
- 4
License it
Wrap the agent binary with Nimbus. Bind the license to the developer's machine fingerprint so the tool follows the hardware, not the user account.
Full source and pre-built binaries available on the dashboard. Requires an active Nimbus license.