datacompose/tobler

Tobler is a tool that decides how much code context to hand an AI coding agent for a given question, instead of always returning a fixed number of results. The idea, borrowed from a geography principle, is that things closer to the query in meaning are more relevant than things farther away, so the amount of context returned should stretch or shrink to match how confident the match is, all within a set token budget. A narrow, specific question like asking what a particular constant does gets back just a few full pieces of code plus a small map. A broad question like how a feature works across the codebase gets back a much wider map of function signatures and summaries. Both answers use the same token budget, just spread differently. The README reports that on large repositories, a fixed top-5 style search misses the right piece of code about two-thirds of the time, while this adaptive approach kept finding it in benchmark tests, using fewer tokens per correct answer along the way. Under the hood it standardizes similarity scores, applies a softmax function to turn them into weights, and uses those weights to decide how many chunks of code deserve attention and how much detail each one gets, from a full code body down to just a file path and symbol name. It also smooths results across a back-and-forth conversation, so a follow-up question like asking about the arguments to something just mentioned stays focused on the same code, while a totally different follow-up question is treated as a fresh topic. It is built on top of another retrieval library called semble and requires Python 3.12 or newer plus a client that supports the MCP protocol, such as Claude Code. It can be installed from PyPI and used either as an MCP server that an AI coding agent talks to directly, or from the command line for scripts and agents that do not support MCP. It also ships as an installable skill that teaches a coding agent when and how to use it. The project is licensed under the MIT License.