Skip to main content
📖 The AI Tool Bible

Vestige

AI diff review for VS Code — see what an AI change actually touches before you merge it

Free· Free (v0.1.0 on the VS Code Marketplace)CodingOpenAI (backend) for labels and summaries; local tree-sitter parsers for structural analysis
Visit website →
Best for

Developers and small teams shipping a lot of AI-generated code in TS/JS, Python, Go or Java who want a structural, low-noise way to review agent diffs inside VS Code before merging.

Skip if

Teams on languages Vestige does not parse yet, strict on-prem shops that cannot send any snippet to a hosted LLM, or anyone needing a mature, production-hardened review platform today.

Vestige is a Visual Studio Code extension aimed at engineers who accept a lot of AI-generated code and need to review it responsibly before it lands on main. Instead of dumping a long unified diff, Vestige reorganizes an AI change into review cards backed by two structural views: a call graph showing what functions the touched code calls into (and is called from), and a file dependency map showing the blast radius across the workspace. The idea is to answer the three questions a reviewer actually has — where do I start reading, what could this break, and what did the surrounding code look like before the model rewrote it — without forcing you to re-derive that context by hand.

The AI role is deliberately narrow. Impact analysis, call graphs and dependency edges are computed locally using tree-sitter WebAssembly parsers, so the structural claims about a diff are deterministic rather than hallucinated. The LLM is only invoked to label nodes and write natural-language summaries of what a change is doing, and only selected snippets — not your whole repo — are sent to the Vestige backend (AWS/OpenAI) for those calls. There is no telemetry.

Supported languages at v0.1.0 are TypeScript/JavaScript, Python, Go and Java, and the extension currently requires a single-root workspace. It is a first public release from an independent publisher, so expect rough edges around incremental parsing and dynamic dispatch, but the core workflow — 'my agent wrote 900 lines, show me the review order and what it touches' — is already usable and slots neatly next to Copilot, Cursor, Claude Code and similar AI coding assistants.

Editor's take

I like the thesis: the interesting problem with AI code is not writing it, it is reviewing it, and Vestige is one of the few tools treating the diff itself as the primary artifact. Keeping the graph deterministic and using the LLM only for prose is the right call. It is very v0.1, so I would use it as a reviewer aid, not a gate — but the direction is exactly where AI-adjacent tooling should be going.

— The AI Tool Bible editorial team

Pros

  • Structural impact analysis runs 100% locally via tree-sitter, so blast-radius claims are deterministic, not LLM guesses
  • Reorganizes large AI diffs into ordered review cards with a suggested reading path instead of a wall of hunks
  • Call graph and file dependency views make it easy to see what a change actually reaches into
  • LLM use is scoped to labels and summaries, and only selected snippets leave the machine
  • No telemetry or analytics — a rare stance for an AI-adjacent dev tool
  • Free on the VS Code Marketplace with no account gate to try it
  • Sits alongside any AI coding assistant (Copilot, Cursor, Claude Code, agents) rather than competing with them

Cons

  • ⚠️ Very early (v0.1.0, ~2 installs at time of writing) — expect bugs and API churn
  • ⚠️ Language coverage limited to TypeScript/JavaScript, Python, Go and Java
  • ⚠️ Single-root workspaces only — monorepos opened as multi-root are not supported yet
  • ⚠️ Acknowledged weaknesses around incremental parsing and dynamic dispatch can hide real edges in the graph
  • ⚠️ Snippet-level summaries still go to a hosted backend (AWS/OpenAI), which some regulated codebases will not allow
  • ⚠️ No pricing/enterprise story published yet, so long-term availability of the free tier is unclear

Use cases

Reviewing large AI-generated pull requestsUnderstanding blast radius of an agent-authored refactorOnboarding to an unfamiliar diff via call graphPrioritizing which files in a big change to read firstSanity-checking a Copilot or Cursor edit before commitExplaining an AI change to a teammate with a generated summaryAuditing what an autonomous coding agent actually touched

Explore related

Compare with similar tools

All in Coding

Cursor

Featured
Coding · Claude / GPT (configurable)
9.5

AI-first VS Code fork — chat, edit, and agentic coding in one IDE.

Freemium· Free hobby; Pro $20/mo; Business $40/mocodingrefactors

GitHub Copilot

Featured
Coding · GPT / Claude / OpenAI o-series (configurable)
9.1

The original AI pair programmer, now with chat and agents.

Paid· Free for individuals; $10/mo Pro; $19/mo Businessautocompletechat

Replit Agent

Featured
Coding · Multi-model (Claude / GPT configurable)
8.7

Build & deploy a full app from a single prompt.

Freemium· Free credits; Core $20/mo; Teams $35/moprototypesinternal tools

Warp

Coding · Multi-model: OpenAI, Anthropic Claude, Amazon Bedrock, plus BYO via OpenRouter and LiteLLM
8.8

The agentic development environment, from the terminal up

Freemium· Free / Build $20 per mo ($18 annual) / Max $200 per mo ($180 annual) / Business $50 per user/mo ($45 annual, up to 25 seats) / Enterprise customAgentic debugging of failing builds and testsNatural-language shell command generation

Cline

Coding · Model-agnostic: Claude (Anthropic), GPT (OpenAI), Gemini (Google), DeepSeek, Grok, Mistral, Cerebras, plus local Ollama/LM Studio
8.7

Open-source agentic coding assistant that plans, edits, and runs code inside your IDE

Freemium· Open Source: free (BYOK or pay model inference at cost via ClinePass) / Enterprise: custom pricing (contact sales for SSO, SLA, RBAC, team management)Multi-file feature scaffoldingLarge-scale refactors

Aider

Coding · BYO (Claude / GPT-4 / Gemini / DeepSeek)
8.4

Terminal-based AI pair programmer that writes commits.

Free· Free / open-source; you pay the underlying LLM API costsCLIgit workflow