Koragraph

The ultimatecontext layerfor AI agents.

Completely free, and it runs locally.

Download on GitHub

github.com/Koragraph/KoragraphMCP

A completely local knowledge graph for your AI agents, coupled with a memory layer that knows when the code it remembers has changed.

Compatible with

Koragraph MCP works with Claude Code, Cursor, Windsurf, Cline, Continue, Codex, OpenCode, Zed, Antigravity, GitHub Copilot.

An agent can hold a fixed amount at once. Without a graph underneath it, that budget fills with whole files, and only three of them hold the answer. Koragraph sends the resolved subgraph instead, so the same answer costs a fraction of the window.

Three repositories are parsed into three graphs, then the calls that leave one service for another are resolved between them into a single graph. Asking about one function returns an answer that has already crossed a repository boundary.

ONE MAP OF EVERYTHING YOU HAVE BUILT.

Koragraph parses every repository you point it at into a single graph, then keeps it current as the code moves. Not an index of text that looked similar to your question, but a resolved structure, with an edge only where there is really an edge.

  1. 01Every service, file, endpoint and table
  2. 02The calls between them, resolved against the syntax tree rather than guessed by name
  3. 03The links that cross a repository boundary, which is where a single-repo tool goes quiet

Research

Our research work.

What we sell, we test first. The design, the task suite and the grading go out with the result, whichever way the result went.

  1. 01

    The Authorship Floor

    Project documentation fails at both ends at once: it never captures what is too small to write down, and it rots what is large enough to write, and both failures share one cause. We describe the memory layer we built on the argument that the two costs have to fall together, and we do not claim it improves outcomes.

    2

    documentation failures, below the floor and above it, from one cause

  2. 02

    94% of Frontier Retrieval Quality from a 494M-Parameter Model

    A 494M-parameter model fine-tuned for 2.43 T4-hours, about $0.85 of commodity GPU time, writes the method descriptions a code knowledge graph is searched on at 93.8% of the frontier teacher’s retrieval quality, and beats a hosted model Koragraph itself was paying for.

    93.8%

    of frontier teacher retrieval quality, hit@8

  3. 03

    Can Graph-Based Context Engineering Substitute for Model Scale?

    A cheap model given graph-derived context solved 9 of 11 repository-level tasks at $0.027 per solved task. A frontier agentic system searching the same repositories with grep solved 9 as well, at $2.82.

    2 → 9

    cheap tier, tasks solved, grep → graph

Read the research →

Korablog.

Notes on what we are seeing in agent research, from the team building Koragraph.

Read Korablog →