Every single session

Explaining your codebase to your agent. Again.

BUILD THE MAP ONCE

Your agent re-reads and re-greps the same files every session to rebuild the context it had yesterday. Nothing persists. justgraphit turns the folder into a knowledge graph once. After that, your agent queries the graph.

Point it at anything.

A path or a GitHub URL. Code, docs, papers, images.

/justgraphit                     # graph the current directory
/justgraphit <path>              # graph a specific path
/justgraphit <github-url>        # clone a repo, then graph it
/justgraphit <path> --update     # re-extract only new or changed files
/justgraphit query "<question>"  # answer from the graph, with citations
/justgraphit path "A" "B"        # how two concepts connect
/justgraphit explain "<name>"    # everything connected to one node

Then the graph becomes the interface. Ask a question and it walks the graph to answer, citing your files. Answers are saved back in, so the graph gets better as you use it. When files change, --update re-extracts only what changed.

Two clusters. One bridge. The red node is the connection you did not think to ask about, the kind community detection surfaces on its own.

Every run leaves three files behind

graph.htmlThe interactive graph. Open it in a browser and wander.

graph.jsonGraphRAG-ready data. The same format upstream graphify writes, so graphs built by either are readable by both.

GRAPH_REPORT.mdPlain language: the god nodes, the surprising connections, and questions worth asking next.

Every edge says how sure it is

A graph you cannot trust is worse than no graph. So every relationship carries a label, and the uncertain ones stay visible.

EXTRACTED

Stated explicitly in your source. The edge cites where.

INFERRED

Read between the lines, and it says so, with a confidence score.

AMBIGUOUS

Uncertain. Flagged for you to judge. Never omitted, never laundered into a fact.

Cost, in the open

Every run reports its token spend, tracked cumulatively in cost.json.

What it won't do

Invent an edgeUncertain relationships are marked AMBIGUOUS, not dressed up as facts.

Hide costBuilding the graph costs LLM tokens. It says how many, every run.

Transcribe videoOut of scope here. Upstream graphify does it.

Answer beyond the graphIf the graph lacks the information, it says so.

Built on Safi Shamsi's graphify

justgraphit is a trimmed packaging of graphify (MIT). Safi's graphifyy Python package is the engine, installed automatically the first time the skill runs, and this skill's pipeline is a derivative of his.

It keeps the core build, update, and query loop and drops the rest. Video transcription, Obsidian, Neo4j, and the MCP server live upstream: for the full surface, use graphify itself.

If it saves you time, support the engine: sponsor Safi on GitHub.

Install

The skill is a folder. No server, no account, no telemetry.

From the skill's directory, one script symlinks it into every Agent Skills directory it finds:

./install.sh

Or do it by hand, into any tool that loads Agent Skills:

ln -s "$PWD" ~/.claude/skills/justgraphit

The graphifyy engine installs itself (uv, pipx, or pip) the first time the skill runs. Python 3.10 or newer.

There is no public repo yet. The skill's profile lives at snitchplugin.com/just/justgraphit. It is part of the just*it family, alongside justloopit.dev.

Questions

Is this free?

Yes. MIT licensed. No account, no email, no plan. Building the graph costs LLM tokens, and every run tells you exactly how many it spent.

Does it work outside Claude Code?

Any AI coding tool that reads Agent Skills. Check your tool’s docs for the Agent Skills format.

How is this different from graphify?

It is graphify, trimmed. Safi Shamsi’s engine does the work; this skill keeps the build, update, and query loop and drops video transcription, Obsidian, Neo4j, and the MCP server. For those, use graphify itself.

Can it read video or audio?

No. Transcription is out of scope here. Upstream graphify does it.