graphify

Graphify is a tool that converts a codebase into a queryable knowledge graph, integrated into AI coding assistants in the form of the /graphify skill. Its officially described goal: by entering /graphify in an AI coding assistant, the entire project (code, documents, PDFs, images, videos) can be mapped into a queryable knowledge graph, rather than searching line by line through files.

Uses

• Convert code, SQL schemas, R scripts, shell scripts, etc. in any folder into visual dependency graphs, architecture diagrams, and ER diagrams.
• Build the codebase, together with documents, SQL schemas, configuration, and PDFs, into a queryable knowledge graph.
• Replace grep and file-by-file reading with graph queries in AI assistants, reducing the agent's token consumption.

Main capabilities

• Code mapping is free and fully local: code is parsed using tree-sitter AST, deterministically, without calling an LLM, and data does not leave the local machine.
• Documents, PDFs, images, and videos are semantically processed using the assistant's model or a configured API key.
• Every edge has an explanation: connections are labeled as EXTRACTED (explicitly present in the source) or INFERRED (inferred by graphify parsing), making it possible to distinguish direct reads from inferred content.
• Not a vector index: no embeddings, no vector store, but a real traversable graph. You can ask questions, trace the path between two things, or explain a concept.
• Output files include graph.html (a clickable graph), GRAPH_REPORT.md (summarizing god nodes, communities, and critical paths for overall architecture review), and persisted graph data (including graph.schema_version and graph.graphify_version).
• Supports local parsing of about 40 programming languages, and can map documents, PDFs, images, and audio/video through optional semantic processing.
• Video/audio is transcribed locally using faster-whisper, and data does not leave the local machine.
• No telemetry, no usage tracking, no analytics.
• Query logging: graphify query, graphify path, graphify explain, and MCP query_graph calls are recorded in JSON Lines format to ~/.cache/graphify-queries.log (timestamp, question, corpus, number of returned nodes,elapsed time); by default, full subgraph responses are not stored.

How to use

• Environment requirements: Python 3.10+ and uv (or pipx).
• After installation, run graphify install once; it detects the assistant you use and registers the skill, with no need for separate configuration per assistant.
• Enter /graphify in the AI assistant (Codex uses $graphify).
• Project-level installation: add --project to write to the current directory, for example .claude/skills/graphify/SKILL.md or .agents/skills/graphify/SKILL.md, and print a git add hint for committable files.
• Platform command examples: graphify claude install --project, graphify codex install --project.
• Generic cross-framework location: --platform agents (alias --platform skills) points to ~/.agents/skills/ (global) or ./.agents/skills/ (with --project).
• Make the assistant always use the graph: run graphify <platform> install once in the project (such as graphify claude install or graphify codex install), which writes a configuration file prompting the assistant to prefer scoped queries like graphify query "<question>" rather than reading the full report or grepping raw files.
• Use the graph directly: in the CLI, query graph.json with graphify query, graphify path, and graphify explain. When a name matches multiple nodes, it refuses and lists the candidates rather than guessing. Scripts can determine the result via exit codes: 0 means there is an answer, 3 means the name is ambiguous, 4 means no matching node.
• Uninstall: graphify uninstall (add --purge to also delete graphify-out/), or uninstall per platform, such as graphify claude uninstall.
• Project configuration: create .graphifyrc in the scan root directory, eachOne key=value per line, lines starting with # are comments. Used to declare ambiguous extensions (e.g. .inc, .h, .m have different meanings in different projects); declarations run through classification, extractor selection, syntax and parsing, imports, graphify watch, and the AST cache.
• Ignore rules: follows .gitignore, supports --exclude patterns and the project .graphifyignore.

Use cases

• Quickly understand architecture, dependencies, and critical paths in large or unfamiliar codebases.
• Let AI coding assistants replace file-by-file reading with graph queries, reducing token consumption.
• Teams share the same graph context and combine it with graph context during PR review.
• For code scenarios with data residency requirements, you can use --backend ollama to run fully locally, or explicitly specify --backend.

Limitations and caveats

• The code portion is fully local and requires no API key; pure code corpora can run graphify extract offline, and mixed repositories can add --code-only to index only code.
• Headless mode graphify extract requires an API key (GEMINI_API_KEY/GOOGLE_API_KEY, MOONSHOT_API_KEY, ANTHROPIC_API_KEY, OPENAI_API_KEY, DEEPSEEK_API_KEY), a running Ollama (OLLAMA_BASE_URL), an AWS credential chain (Bedrock, no API key needed, uses IAM), or the claude CLI binary (Claude Code, no API key needed, uses a Claude subscription) when processing documents, PDFs, and images.
• Data residency: graphify extract automatically detects the provider based on the API keys that are set (priority: Gemini → Kimi → Claude → OpenAI → DeepSeek → Azure → Bedrock → Ollama); Kimi (MOONSHOT_API_KEY) routes to Moonshot AI servers in China.
• Some platforms require additional configuration: Codex requires, under [features] in ~/.codex/config.toml,setting multi_agent = true to extract in parallel; Pi requires version 0.58.4 or newer and must be persistently installed (uv tool install graphifyy or pipx install graphifyy, not uvx).
• The official PyPI package name is graphifyy (double y); other graphify* packages are unrelated to it; the CLI command is still graphify. The official source repository is Graphify-Labs/graphify.
• The official Graphify Cloud 14-day free trial is offered for continuous operation, single-repository and cross-repository support, formal verification, SDLC-scope code review, and connectors such as Sentry and Jira; specific billing and pricing information is not yet provided.
• Benchmark data: 0.497 recall@10, 0 LLM credits; the full table and reproduction commands are in BENCHMARKS.md.
• The full guide and reference are at docs.graphify.com.