Open-source Graft gives coding agents a map of your codebase, claiming up to 4x savings

A new open-source tool called Graft is getting attention from the developer community for a simple pitch: it gives your coding agent a map of your codebase. Graft builds a structured context graph for large codebases so that coding agents such as Claude Code, Cursor, Codex, and Gemini can understand the repository's structure and dependencies before touching a line of code.

Graft comes from enterprise AI company NanoNets and is hosted on GitHub at NanoNets/Graft, built in TypeScript and released under the MIT license. Setup is minimal: install the @nanonets/graft package via npm, run graft init, and the tool builds the graph and wires it into workflows like Claude Code.

Under the hood, graft init constructs a graft/ graph directory from the code and pulls matching nodes into each prompt, giving the agent relevant structural context at every step; the graph is rebuilt incrementally to avoid full rescans of large repositories. Commands such as graft grep, graft map, and graft viz let developers search, orient, and visualize the codebase structure.

Efficiency is the headline claim. In a controlled 162-run benchmark — same agent, same file tools, only the context differs — Graft reports a 46% reduction in tool calls, 42% token savings, and 60% time savings; on SWE-bench Verified, the correctness of Claude Code with Graft rose from 54% to 66%. The project advertises the result as 'up to 4x cheaper and 3x faster.'

Why it matters: context is the biggest bottleneck for coding agents today. The larger the repository, the more tokens agents waste on irrelevant files — and the easier it is to edit the wrong place. By offering a codebase map as infrastructure that works across mainstream agents, Graft attacks cost, speed, and correctness with a single graph, and its open-source release means developers can audit and self-host it.

The broader signal is competitive: as Claude Code, Cursor, Codex, and Gemini spread quickly, the middleware layer for 'helping agents understand your repo' is becoming a hot new category, with MCP servers, code graphs, and semantic retrieval all competing for the same slot. Graft deliberately supports multiple agents rather than locking into one product.

What to watch next: whether Graft's benchmark numbers hold up across real repositories, how it performs on very large monorepos, and whether context-layer tools become standard equipment for coding agents — making 'build the map, then code' a default developer habit, much like IDE indexing.