Graphify Explained

Graphify (Graphify-Labs/graphify) is an open-source Python tool that turns a codebase, plus its docs, SQL schemas, configs and PDFs, into a queryable knowledge graph that coding agents such as Claude Code, Cursor and Codex can query through a skill, a CLI or MCP. The verdict: it is a graph-based alternative to vector search for the problem of agents re-grepping a large repo and misreading its structure, and the practical point is that code parsing runs locally with no LLM.

The repository sits at roughly 135k GitHub stars as of October 2026, which makes it one of the most-watched agent tooling projects. Everything below comes from the official README: how it works, install, pricing, and how it differs from the alternatives.

Graphify parses source code with tree-sitter and turns functions, classes, calls and imports into nodes and edges. Docs, PDFs and images go through a semantic pass that uses your agent's model (or a configured API key) and are merged into the same graph. Each edge is tagged EXTRACTED (explicit in the source) or INFERRED (resolved by Graphify), so you can tell how strong the evidence is.

Note that the PyPI package name is graphifyy (two y's). PDF, Office, video, MCP, SQL and Leiden support come as extras (pdf, office, video, mcp, sql, leiden, all). The core is free; the LLM-based semantic pass costs whatever your agent's model or API key costs. The hosted Graphify Cloud advertises an always-on graph, monorepo and cross-repo support, code review and Sentry/Jira connectors, and the README claims an 80% cut in token spend. We found no independent verification, so treat that number with caution.

- 1. Install the CLI with uv tool install graphifyy, then run graphify install to register the skill with your agent
- 2. Open your agent at the project root and run /graphify . to build the graph ($graphify in Codex)
- 3. Read graphify-out/GRAPH_REPORT.md and graph.html for the overall picture and communities
- 4. From then on, query with /graphify query, path and explain; after code changes, run --update to refresh only the diff

It helps to think in layers. Graphify is the structural map of the codebase, ai-memory and Hindsight handle work memory and handoff, OpenViking holds the broader context, and Serena does precise symbol-level lookup and edits. ai-memory does not analyze code relationships; it is a separate layer for session handoff, so it complements Graphify rather than competing with it. For a small repo, Serena or plain grep is often enough; Graphify pays off once you need to follow cross-cutting dependencies and design intent.

- The semantic pass for docs, PDFs and images uses an LLM, so check where content goes (your agent's model or API key) before feeding in confidential documents. Use GRAPHIFY_BACKEND=ollama or none to avoid external calls
- The "80% fewer tokens" figure is a Graphify Cloud claim with no third-party verification. Measure it on your own repo
- The published benchmarks (LOCOMO recall@10 0.497, LongMemEval-S QA 76%, etc.) are mostly memory and QA metrics, not a guarantee of better coding results
- INFERRED edges can be wrong. Verify important decisions against EXTRACTED evidence or the actual code
- With this many stars come many open issues and PRs (several hundred each as of October 2026), so commands and behavior may change often. Pin the version
- Query logs are written to ~/.cache/graphify-queries.log; set GRAPHIFY_QUERY_LOG_DISABLE=1 to turn that off
- For projects of a few dozen files, building a graph is usually not worth the effort

For related reading on agent context and memory, see OpenViking (a context database for agents), Hindsight (agent memory) and archify (drawing architecture diagrams with an agent skill); together they make the division of labor between these tools easier to see.