Local MCP server that gives AI agents semantic cross-repo code context
kivgraph, created by Luqueee, is an MCP server that gives AI coding agents a persistent, language-aware view of large codebases. It builds a cross-repository semantic graph and exposes query tools so agents can trace dependencies, find declarations, and assess change impact with fewer tokens. Key elements include compiler-driven edge resolution, persistent local storage, and a 3D read-only explorer. The tool targets engineers and teams using AI assistants who need precise cross-repo reasoning and reduced token overhead.
What tasks can you actually use it for?
The tool exposes a set of MCP actions that let agents query relationships directly instead of rereading source files. Notable operations named in the project include find_references, trace_dependencies, and find_by_intent. Typical agent tasks are dependency tracing across repositories, locating symbol declarations, and calculating change impact so automated edits can be scoped. Results return structured relationships rather than raw text matches, enabling targeted reasoning by downstream agents.
How accurate and token-efficient are the outputs?
Benchmarks published by the project report 7.4x token efficiency versus grep-and-read approaches for repository navigation. The tool resolves edges using language-native analyzers and compilers, so relationships carry explicit compiler evidence that the developer describes as the source of each edge. Semantic indexing currently covers Go, TypeScript, Rust, Python, and Dart, which narrows the domains where that accuracy applies today.
Is it straightforward to install and integrate into developer workflows?
The tool installs as a self-contained bundle through a single-command shell script and requires Node.js 22 plus standard shell utilities. It runs as an MCP server compatible with MCP-enabled clients such as Claude Desktop, Cursor, and Claude Code. Published binaries target Linux amd64 and macOS Apple Silicon, so teams on those platforms can add the server to existing agent pipelines without separate cloud indexing services.
Does it protect code privacy and who maintains the project?
The tool operates locally with no API key dependency, so indexing and queries do not require external indexing services. The developer publishes the source under Apache License 2.0 on GitHub and keeps the project actively maintained. Community discussion and repository benchmarks referenced by the project show positive reception around token savings and navigation accuracy, supporting adoption by teams that prefer open-source, locally run tooling.
The tool suits teams that need precise AI-driven code context
The tool is a pragmatic choice for engineering teams that integrate MCP-enabled assistants and need programmatic answers rather than rereading many files. It requires operator familiarity with a query-based workflow, so budget time for onboarding and validation. Practical advice: use targeted semantic queries and follow up with human code review before applying automated edits, keeping risk low during large-scale refactors.
