Anytime Graph

日本語 | English

See what gets talked about together — as a diagram, inside VS Code.

Interview notes, review comments, and free-text survey answers hide their structure: a flat list never shows which words keep appearing together. Moving the text into an external analysis tool works, but every revision costs another round trip.

Anytime Graph edits co-occurrence networks (*.cooc.json) in a VS Code custom editor, so you can let an AI generate the network and then refine it by hand without leaving the editor.

Try the online editor

1. Features

• Visual co-occurrence editing — circle size = frequency, line = co-occurrence strength, color = cluster
• Term editing from a panel — search, add, rename, set frequency, assign a cluster, delete
• Filters that keep only what matters — thin the view by minimum frequency, minimum strength, or top-N links
• Force layout runs in a Worker — the UI stays responsive on large networks, and you can abort mid-run
• Positions cached in the file — reopening keeps the same arrangement; it recomputes only when the content changes
• PNG export — drop the diagram straight into a document
• Sidebar list — browse and open every *.cooc.json in the workspace

2. Getting Started

1. Open Anytime Graph from the activity bar
2. Click + in the Co-occurrence Networks view title bar (Anytime Graph: New Network)
3. Enter a file name (defaults to untitled.cooc.json)
4. An empty network is created and opened in the custom editor

Existing *.cooc.json files open in the custom editor straight from the explorer.
New files are created directly under the first workspace folder (names containing a path separator are rejected).

3. Editing Terms and Co-occurrences

Terms are edited from the Words panel on the right. Clicking a term in the diagram selects it in the list as well.

There is no GUI yet for adding co-occurrences (the lines between terms) or changing their strength. Build networks with links via "6. Generating from an AI (MCP)" below, or edit the file directly.

4. Filtering and Display

Large networks become unreadable. Use the Filters panel on the left to thin the view.

• Minimum frequency — hide terms below the threshold
• Minimum strength — hide weak co-occurrence lines
• Top links — keep only the N strongest co-occurrences
• Filters change the view only; the file content is untouched
• Hidden terms are marked as hidden from the diagram by the current filter in the list

The toolbar offers Fit (fit the whole diagram on screen), PNG (export), Save, and Show/Hide panels.
Display language and color theme follow your VS Code settings.

5. Layout

Positions come from a force layout (repulsion, attraction, and gravity toward the origin). It runs in a Web Worker, so the editor stays usable while it computes, and the toolbar Abort stops it.

• Results are cached as coordinates under layout in the *.cooc.json file
• The cache is validated against a content hash and an algorithm version, so it recomputes only when terms or co-occurrences change
• Isolated terms with no co-occurrences are pulled toward the origin instead of drifting off-canvas

6. Generating from an AI (MCP)

The MCP server (mcp-graph) ships with the extension — nothing extra to install. It writes a co-occurrence network straight out of a text analysis. Endpoints are given as term labels, so there is no need to count indices.

Wiring happens automatically.

• VS Code (Copilot / Chat) — the extension provides the server; no configuration needed
• Claude Code — on activation the extension adds mcp-graph to the workspace .mcp.json. An existing entry of the same name is never overwritten (so a hand-tuned setup survives). To rebuild the entry, run Anytime Graph: Register MCP Server in .mcp.json from the command palette

The intended split is: generate the file with the MCP tools, then open it here and refine frequencies and clusters by hand.

7. File Format

Co-occurrence networks are stored as *.cooc.json — plain JSON, friendly to version control.

{
  "meta": { "schemaVersion": 1, "generatedAt": "2026-07-22T00:00:00.000Z", "origin": "manual" },
  "spec": {
    "title": "Interview analysis",
    "subject": 0,
    "nodes": [{ "label": "deadline", "frequency": 12 }],
    "links": [[0, 1, 0.8]],
    "clusters": [{ "label": "constraints", "members": [0, 1] }]
  }
}

• links are [term index, term index, strength]. A term's index in nodes is its identity, so reordering nodes or deleting from the middle by hand changes what existing links point at
• The term named by subject is emphasized with a thicker outline
• layout is only a position cache and can be omitted when writing the file by hand

8. Commands

9. License

MIT