Semantic Vision

AI can write your code in seconds. Understand it just as quickly — right inside VS Code.

Semantic Vision is a codebase observability layer for AI-assisted coding: a map of the codebase that you and your AI agents can both use. Impact analysis, an interactive call graph, execution flowcharts, seven-signal Code Health, AI-generated documentation, and code-to-data lineage — the same engine as the Semantic Vision web app, as a panel next to your code instead of a separate browser tab.

Why

Whether a teammate wrote it or an AI agent did, code still has to be understood before you can trust it. The traditional way — grep, jump between files, trace dependencies by hand, build a mental model, hope you didn't miss something — takes as long as it ever did, no matter how fast the code got written. Semantic Vision replaces all of that with one question: right-click a function and see its blast radius, its behavior, or its health, in seconds instead of an afternoon.

Impact analysis

Right-click any function — or, with code-to-data lineage connected, any table or column — and choose Impact Analysis to see every direct and transitive caller, resolved from the real call graph, not a guess. Circular call chains are flagged instead of silently mishandled, and the whole blast radius highlights live on the graph. This is the answer to "what actually depends on this" before you merge a change, not after something breaks in production.

• Semantic Vision: Open Graph (semanticVision.openGraph) — opens the full call graph in a panel, centered on whichever file is active in the editor. Click any node to jump straight to that function or class in your code.
• Semantic Vision: Impact Analysis at Cursor (semanticVision.impactAnalysisAtCursor, also on the editor's right-click menu) — resolves your cursor position to a graph node and highlights its full blast radius without leaving your place in the file.

Everything else, in the same panel

Once the graph is open, every other feature works exactly as it does in the web app.

Interactive call graph

Every directory, file, class, and function as a zoomable, color-coded graph with call/import/defines edges — structure that would otherwise take an hour of grepping to piece together by hand. Real-time search across the whole repo or scoped to one file, and your dragged layout persists automatically.

Execution flowcharts

Right-click any function for its execution flowchart, built from its real AST/CST: entry and return points, decisions with Yes/No edges, loops with a visible back-edge, I/O calls, and calls out to other functions — each in its own conventional shape. Works for Python, JS/TS, and Java, including switch fallthrough, do...while's bottom-condition check, and labeled break/continue.

Code Health

A peer lens to the codebase graph, switched via a header tab: a filterable, sortable ranked list across seven signals — Complexity (a real AST walk, not a line-count guess), Hotspots (complexity × how often a file actually changes), Dead code (zero-caller candidates), Coverage risk (complexity × blast radius × untested, from a coverage.py or lcov report you provide), Duplicates (identical once names and literals are stripped), Dependencies (known vulnerabilities via osv.dev — opt-in, the one signal that reaches the network), and Recommendations (an AI pass prioritizing across every signal above at once).

Click any row to see exactly who calls it and what it calls, live, or diff against any commit to see whether an edit helped or hurt.

AI-generated documentation

Right-click any function and choose Document to stream real Markdown documentation (Purpose, Parameters, Returns, Side Effects, Notes) assembled from its actual source, callers, callees, and parent class. Right-click a file for a module-level summary. Pick a local Ollama model (free, nothing leaves your machine), OpenAI, or Anthropic — nothing is saved until you click Save.

Code-to-data lineage

SQLAlchemy models are detected automatically on every parse; connect a dbt manifest.json and/or a live database connection string to add their tables, models, and columns to the same graph, reconciled by name. Flip Data only to read it as a pure lineage diagram, with impact analysis spanning code and data in one traversal.

Your AI agent can see the map too

The same analysis this extension shows you is also available to coding agents over MCP: call graphs, impact analysis, complexity, hotspots, dead code, coverage risk, duplicates, execution flow, and more, as tools Claude Code, Cursor, or any MCP-compatible agent can call directly — instead of grepping for context this project has already computed. Run semantic-vision-mcp from the main project alongside this extension; see guides/mcp-server.md for setup.

Requirements

This extension is a thin client: it renders the same frontend as the web app, backed by a FastAPI server that does the actual parsing.

• Zero setup (default) — on Windows, macOS (Intel and Apple Silicon), and Linux, the backend ships bundled inside the extension and starts automatically the first time you run a command. No Python, no uv, nothing beyond VS Code itself.
• semanticVision.backendUrl (default http://localhost:8000) — where the backend is expected to answer. If something is already running there (e.g. via docker compose up, see the main project's README), the extension uses it as-is instead of starting its own.
• semanticVision.backendPath — optionally, an absolute path to a local Semantic Vision checkout (the directory containing pyproject.toml). When set, the extension starts the backend from source with uv run uvicorn semantic_vision.api.app:app instead of the bundled binary — useful for contributors working against an unreleased backend change. Requires uv on PATH and a prior uv sync in that checkout.

Nothing about your code is ever executed or sent anywhere — parsing is static, and the backend runs entirely on your own machine. Nothing leaves it unless you explicitly ask for AI-generated docs from a cloud provider, or opt in to a dependency vulnerability scan.

Getting started

1. Install this extension.
2. Open a file in a supported repository (Python, JavaScript/TypeScript, or Java).
3. Run Semantic Vision: Open Graph from the Command Palette, or right-click a function and choose Semantic Vision: Impact Analysis at Cursor. The bundled backend starts automatically — no extra setup.

Learn more

Full documentation, screenshots, and setup guides (AI provider setup, the MCP server, code-to-data lineage, multi-language support) live in the main project's README and guides/ directory.