Sunstone
Stop your coding agent guessing about your own codebase
Start new thread

Sunstone - Stop your coding agent guessing about your own codebase

by•
Your coding agent already has your files. It greps, misses, and answers anyway. Sunstone indexes the folders you open and gives whichever model you use, Copilot's included, tools to search them properly: passages with the file and function they sit in, whole-file digests, lookups with citations, notes that survive chats. On SWE-bench Verified it names the right file first 229 of 500. Text search gets 58. Your own servers can join the picker too. It stays on your machine. Free. Source public.

Add a comment

Replies

Best
Maker
📌
Hi Product Hunt. Sunstone started because we kept watching very good models be confidently wrong about code that was sitting on disk three directories away. We measured it rather than guessed. Claude Sonnet 5 and Claude Opus 5, given every file of our own published figures plus grep and cat, asked what those figures say. Across 56 questions and four runs each, one answer in five came back wrong. Nothing was missing. Finding it was the problem, and grep is a bad way to find a thing you can only describe. Then we changed one thing, not the model and not the prompt. The same files became a searchable store, and the models got tools to query it. The scores went from 79% and 80% to 93% and 95%, every run with the store beat every run without it, and the work finished around three times faster while reading a quarter as much. That is 56 questions over a claims register, so it is a document-retrieval result rather than a coding benchmark, and I would rather say so than let it imply more. For code specifically, the retriever names the correct file first for 229 of 500 bug reports on SWE-bench Verified. An ordinary text search gets 58, and the same queries pointed at the wrong repository land 5. Every ranked list is published. Sunstone works with whatever model you are already using, Copilot's Claude and GPT included, because the tools are registered where any model in the picker can call them. If you would rather run your own, llama.cpp or vLLM appears beside the hosted models with tool calling, and a context window too small for the conversation gets folded into the index rather than truncated. The index and the text stay on your machine. No account, no key, and the extension host opens no outbound connection of its own. That last one is checkable instead of promised, which is why the source is public. Free. I would most like to hear where the retrieval fails on your repository, because that is the part I can actually fix.