Developers can map code structure and documentation into a local knowledge graph for coding agents
When you need to locate a specific feature in an unfamiliar codebase or trace how a single line change impacts a service, you can find answers without opening every file. This tool bundles source code, documentation, and infrastructure settings into a single persistent graph managed locally. Coding agents use this graph to gain structural context, while developers can directly explore code flow using a visualization tool.
A code knowledge graph running locally
The repository operates around an MCP (Model Context Protocol) server and uses a native backend written in C. It performs indexing and graph queries without external hosting services or API keys, so all processing happens on the user's machine. It exposes 17 MCP tools and parses various languages and file formats using vendored grammars. Supported languages include hybrid LSP type interpretation for more accurate analysis. The index persists across sessions, reducing repetitive file reads and token usage.
CBM Lens: A visualization tool for 3D code structure
CBM Lens is an optional visualization companion currently in open beta (v0.0.1). Folders appear as nested platforms, files and symbols as selectable blocks, and indexed relationships as arrows. In the 3D galaxy view, you can explore symbol exit relationships by layer and switch to a hierarchy view to trace call paths. The source reader displays syntax-highlighted code in a Monaco editor, allowing you to immediately check the callers, callees, and related tests of the symbol under the cursor. You can also ask questions about selected code using a local model in the browser, with all inference performed on the device.
Agent integration and session coordination
The installation command automatically configures settings, skills, agent profiles, and hooks for the selected client. Scout, Verify, and Auditor profiles provide quick temporary lookups, task-oriented verification, and limited full verification, respectively. A session coordination daemon shares a single daemon across multiple clients like Claude Code, Codex, and OpenCode to manage watchers, indexing jobs, and the UI. All active processes must use the same version and cache root; if a conflict occurs, it fails before the job starts and logs the error. During updates, the daemon and all temporary jobs are safely terminated before activating the new binary.
Tracking connections to infrastructure and documentation
It tracks relationships beyond source files, extending to infrastructure configurations like Docker, Kubernetes, Terraform, and Ansible, as well as CI settings. Documentation sections, PDF text pages, and Architecture Decision Records (ADRs) are linked to the code, so explanations are found right next to the symbols. The experimental Smart CI feature combines static relationships and per-test coverage to select tests affected by changes. If evidence is insufficient, it falls back to a broader execution scope. These features help identify what is affected before executing changes.
Pre-use checks and limitations
Because the tool assumes local processing, it works without a network connection, but using browser AI features requires WebGPU support. The local model in CBM Lens downloads a default 0.5B model of approximately 567MB, with larger models requiring additional capacity. Microsoft Defender may falsely flag the binary as malware, which is a known false positive also reported for other major tools. The ADR editor does not provide atomic document locking, so you must verify changes before saving. Additionally, the tool does not edit code or execute tools, and text generated by models must be verified against the source.
