Scribe

Scribe is a command-line tool that automatically builds a searchable knowledge base from development work without requiring manual note-taking or documentation effort. It solves a persistent problem for developers who use AI coding assistants like Claude Code or Codex: the insights, decisions, debugging solutions, and architectural explanations generated during coding sessions disappear once the conversation ends. Scribe captures this information from git history, AI coding sessions, and self-sent URLs, then transforms it into organized wiki pages that persist across projects and remain queryable long after the original work is done. The accumulated knowledge compounds over time, making past solutions accessible when similar problems arise in future projects.

The tool runs as a single binary built in Go, available for macOS and Linux under the MIT license. Installation happens through Homebrew or a shell script, and once configured, Scribe operates autonomously on a cron schedule with no ongoing intervention required. Users set it up once, configure their input sources, and let it run in the background continuously absorbing new information from their development workflow. The capture pipeline pulls from four input streams: git repositories where commit messages and code changes tell a story of decisions made, AI coding assistant sessions where problem-solving conversations contain valuable context, user-submitted URLs pointing to documentation or resources worth remembering, and iMessage self-chat links where developers often save useful articles or tools they want to revisit later.

Processing happens in stages designed to filter noise before anything reaches the language model, which keeps costs down and quality up. First, FTS5 keyword-density scoring removes boilerplate content, template code, and repetitive patterns that would waste inference time without adding knowledge value. Then a two-pass absorption process creates entity-first wiki pages, meaning the system organizes knowledge around specific concepts, tools, patterns, or decisions rather than dumping everything chronologically. The output is a typed knowledge graph stored as plain markdown files with wikilinks connecting related concepts and backlinks showing which pages reference any given topic. This structure makes navigation intuitive and helps users discover connections between pieces of information they might not have linked mentally.

The graph structure uses ten defined edge types to represent how different pieces of knowledge relate to each other. These include relationships like supersedes (when a newer approach replaces an older one), contradicts (when information conflicts between contexts), specializes (when one concept is a more specific version of another), and several others that capture the nuances of how knowledge evolves. When a debugging approach that worked in one project gets refined or replaced in another, those relationships are explicit in the graph structure rather than buried in chronological notes. This makes it possible to understand not just what was learned but how that understanding developed over time and which information should take precedence.

One of the more practical features involves agent integration with AI coding assistants. Scribe writes handshakes into Claude and Codex configurations so that these AI assistants automatically query the knowledge base before making suggestions or decisions. This means solutions discovered in one project can inform work in another without the developer needing to remember relevant context and re-explain it. The accumulated knowledge becomes part of the information available to AI assistants, reducing repetitive explanations and improving the quality of suggestions. For developers who work across multiple codebases, this cross-project memory eliminates the frustration of solving the same problem twice because the first solution was in a different repository's chat history.

The system can run entirely locally using Ollama for inference, eliminating API costs for users who prefer to keep everything on their own hardware and maintain complete privacy over their development knowledge. Those who want faster processing or higher-quality inference can connect to hosted providers at roughly fifty-five cents to six dollars per month, or use their existing Anthropic Claude CLI subscription. This flexibility means developers can choose their preferred tradeoff between cost, speed, privacy, and inference quality based on their specific needs and constraints without being locked into a particular approach.

Target users are developers working across multiple projects who rely on AI coding assistants and find themselves repeatedly re-explaining context or rediscovering solutions they already found. The tool particularly benefits anyone frustrated by the ephemeral nature of AI conversations, where valuable debugging sessions, architectural discussions, and implementation explanations vanish after the chat window closes. Storage in a plain git repository ensures no vendor lock-in, and the markdown format means the knowledge base remains readable, portable, and accessible regardless of whether Scribe continues to exist. The entire project is free and open source under the MIT license.