Graphify ships v0.9.28 and relicenses to Apache-2.0

Most AI coding assistants retrieve context the same way: embeddings, similarity search, or raw file reads. Graphify takes a different position. No embeddings, no vector store; a deterministic AST graph you traverse with explicit, explained edges. The bet is that structured retrieval with confidence tags beats probabilistic matches when you need to know why two symbols are connected, not just that they are.

At 97.1k stars and 174 releases, the project is past early traction. This week's release adds a stricter enforcement mode for the Claude Code integration and closes out a significant licensing move.

What Happened

The Graphify-Labs/graphify repository shipped v0.9.28, its 174th release, with 97.1k stars, 9.4k forks, and 192 contributors. The project comes from Y Combinator S26 company Graphify-Labs, under an Apache-2.0 license since July 22, relicensed from MIT.

Install is two commands: uv tool install graphifyy (the PyPI package carries a double-y), then graphify install to register the skill with your assistant. Run /graphify . and you get three files: graph.html, GRAPH_REPORT.md, and a graph.json you query without re-reading source.

Code parses locally through tree-sitter across roughly 40 languages with zero LLM calls. Docs, PDFs, images, and video route through your IDE session's model only if you configure one. A code-only corpus runs fully offline.

Key Features

• Strict mode for Claude Code: graphify install --project --strict blocks the first raw source read of a session and redirects it to the graph, then reverts to a soft nudge. Fires at most once per session; never gets stuck. Toggle at runtime with GRAPHIFY_HOOK_STRICT=1 or 0. The default install remains a nudge, not a block.
• Query logging now opt-in: Set GRAPHIFY_QUERY_LOG_ENABLE=1 to enable the local query log at ~/.cache/graphify-queries.log. Previous versions logged by default; if you were relying on the log, update your setup.
• --code-only flag for headless extract: graphify extract ./src --code-only indexes just the code via local AST, skips docs, PDFs, and images, and requires no API key.
• --no-gitignore flag: Pass --no-gitignore to graphify extract when git-ignored generated or transpiled code belongs in the graph. .graphifyignore still applies.
• --allow-partial flag: When an extraction pass crashes on a large corpus, graphify now refuses to overwrite a larger existing graph with a partial result by default. Pass --allow-partial to override.
• Every edge explained: Each connection carries EXTRACTED (explicit in source) or INFERRED (resolved by graphify), so agents and developers can tell what was read from what was guessed.

Why It Matters

The --strict mode is the change most teams running Claude Code should evaluate. The default install nudges Claude toward graph queries before file reads; strict mode enforces it for the first read of a session. For large codebases where an agent opening files sequentially is a real cost, that difference matters. The opt-out is a single env var, so the risk of it breaking a workflow is low.

The Apache-2.0 relicense matters for teams evaluating whether to embed graphify in internal tooling or redistribute it. Apache-2.0 adds an explicit patent license and clearer attribution requirements compared to MIT, which is generally more permissive for enterprise use. If your legal team previously ruled out MIT for internal tooling, worth re-evaluating.

The query logging change is a breaking behavior shift. Teams that built monitoring or audit workflows on top of the log need to add GRAPHIFY_QUERY_LOG_ENABLE=1 to their environment before upgrading.

Example Use Case

The README runs graphify on the FastAPI codebase. After /graphify ., asking graphify path "FastAPI" "ModelField" returns the shortest path in three hops with zero files opened: FastAPI --uses--> DefaultPlaceholder <--references-- get_request_handler() --references--> ModelField.

For a team onboarding onto a large codebase, the workflow is: one developer runs /graphify ., commits graphify-out/ to the repo, and every subsequent developer's assistant reads the map on first checkout. Run graphify hook install to rebuild the graph on every commit via AST only, no API cost. Two developers committing in parallel never get conflict markers in graph.json because the hook sets a union merge driver automatically.

With --strict enabled in Claude Code, new team members' agents can't work around the graph on their first session; they have to go through it.

Competitive Context

Graphify plugs into Claude Code, Cursor, Codex, and 15+ other assistants rather than competing with them. On Claude Code, graphify claude install adds a PreToolUse hook that fires before search and file-read calls. On Cursor, it writes .cursor/rules/graphify.mdc with alwaysApply: true since Cursor doesn't support hooks the same way. On Codex, AGENTS.md carries the always-on guidance since Codex Desktop rejects hook-based nudges on PreToolUse.

The direct benchmark targets are mem0 and supermemory. The project's own BENCHMARKS.md reports recall@10 of 0.497 versus mem0's 0.048 and supermemory's 0.149 on LOCOMO (n=300), and 76% QA accuracy on LongMemEval-S tied with dense RAG. Those are self-published numbers on Graphify's own harness, with judge agreement blind-validated at 90.6%. Read them as directional until independent reproduction appears.

My Take

The --strict mode is the most practically significant change in this release for teams actively using Claude Code. The soft nudge was easy for an agent to work around; strict mode closes that off for the first read of a session. Whether that matters in practice depends on how often your agents were actually bypassing the graph when the nudge was soft, which is now something you can measure by comparing session costs before and after.

The relicense deserves more attention than release notes usually give it. Apache-2.0 at 97k stars means a lot of downstream tools and enterprise setups need to review their dependency posture. The project moved in the right direction for broad adoption, but it's a change worth flagging to your legal team if graphify is already in your stack.

[ Meet Cosmos ]

Run your software agents at scale

Cosmos gives your agents the context, tools, and feedback loops they need to get better with every workflow.

Try it out