context-explorer

See what Claude Code actually loads into its context when it starts in a project — as an interactive graph.

Point it at any folder. If that folder is a Claude Code project (it has a CLAUDE.md, AGENTS.md, or a .claude/ directory), context-explorer discovers everything Claude Code injects at launch — instructions, memory, skills, subagents, slash commands, MCP servers, settings — plus the cross-references between those files, and renders a single self-contained HTML graph you can open offline.

It answers a question that's otherwise invisible: given this repo, what does the model see before it reads a single line of code, and where did each piece come from?

launch-context.svg — the specification of what's injected at agent launch, by level and load type (this is the static model; the tool renders the live graph for a specific project).

Why

Claude Code assembles its launch context from several levels — your user-global ~/.claude, the project's .claude/, and the per-project memory it maintains. Skills and subagents are injected as name + description; CLAUDE.md / AGENTS.md are injected in full; memory findings and context files are pulled in later, on demand. That structure is real but normally lives only in your head. This tool draws it.

Usage

No dependencies — Python 3 standard library only.

python3 context_graph.py /path/to/project          # build -> /path/to/project/context-graph.html
python3 context_graph.py .            --open        # current dir, open in browser
python3 context_graph.py /path/to/project -o graph.html
python3 context_graph.py /path/to/project --no-global   # project level only, ignore ~/.claude

Then open the generated context-graph.html in any browser (works via file://, fully offline).

What it shows

Levels (each is its own anchored cluster):

• User-global — ~/.claude/ (CLAUDE.md, skills, subagents, commands, settings): applies to every project on the machine.
• Project — this repo's .claude/, CLAUDE.md/AGENTS.md, .mcp.json.
• Auto-memory — ~/.claude/projects/<project>/memory/: the per-project memory Claude maintains.

Two kinds of edges:

• Load edges (solid gold) — what a level injects into context at launch. These connect each level's anchor to its CLAUDE.md/AGENTS.md (full content), skills & subagents (name + description), MCP servers and settings (config), and follow @path imports recursively (Claude Code expands those into the prompt at startup too).
• Reference edges (faint, dotted) — textual cross-links between files: markdown links, [[wikilinks]], and inline path mentions.

Each node carries a badge telling you whether it's loaded at launch (full / description-only / config) or on demand (grep / recall).

Context depth (the slider, bottom-right) is the main lens. It's hop-distance from the launch-loaded core:

• 0 — only what's injected when the agent starts: the level anchors plus CLAUDE.md/AGENTS.md, skills, subagents, MCP, settings, and any @imports. Nothing the agent would have to go fetch.
• 1, 2, … — adds everything reachable in N reference hops outward (e.g. the findings that MEMORY.md indexes appear at depth 1).
• all — the complete graph, including files unreachable from the launch core.

Toggle a level off in the legend to isolate the others.

Click any node to read its rendered markdown (or, for .json configs, the redacted contents). You can edit in the panel; edits persist to your browser's local storage and can be downloaded — the standalone file never writes to your disk.

Privacy

• The generated HTML embeds the text of the files it graphs — treat it like the files themselves and don't commit it (the repo's .gitignore excludes *.html).
• settings.json / settings.local.json / .mcp.json are redacted before embedding: values under key/token/secret/password-like keys, and long token-shaped strings, become ***REDACTED***.
• .env files are never read.
• Nothing is sent anywhere; the tool only reads files and writes one HTML file.

License

MIT — see LICENSE.