Vexp — Local-First Context Engine for AI Coding Agents

TL;DR

TL;DR: Vexp is a local-first context engine that builds a dependency graph from your codebase’s AST and serves only the relevant subgraph to AI coding agents — cutting token usage by 65-70% with no data leaving your machine.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: vexp.dev
• Source repository: github.com/Vexp-ai
• License: proprietary (free tier available, paid plans from $19/mo)
• HN launch thread: news.ycombinator.com/item?id=47136384
• VS Code Extension: marketplace.visualstudio.com/items?itemName=Vexp.vexp-vscode
• npm package: www.npmjs.com/package/vexp-cli

What Is Vexp?

AI coding agents like Claude Code and Cursor read entire files linearly to build context. On a medium TypeScript project, a single query can consume ~18k tokens — most of it irrelevant. Vexp solves this by building a dependency graph from the AST (who calls what, who imports what, what types flow where) and serving only the relevant subgraph as a token-budgeted capsule.

Key claims from the project:

• 65-70% token reduction compared to feeding entire files
• Zero network calls — everything runs locally on your machine
• No account required — no cloud, no sign-up, code never leaves your laptop
• AST-based dependency graph — understands call graphs, import graphs, type flows
• Single-call pipeline — serves a curated context capsule in one shot
• Works with Claude Code, Cursor, Windsurf, Codex, and Opencode

How It Works

Vexp sits between your codebase and your AI agent. When you run a query, instead of dumping entire files, it:

1. Parses the relevant source files into an AST
2. Builds a dependency graph (calls, imports, type flows)
3. Identifies pivot nodes — the functions/types closest to your query
4. Expands outward along the dependency edges, respecting a token budget
5. Returns a skeleton context — only the relevant subgraph

The result is a lean, targeted context capsule instead of a 10,000-token file dump.

Setup

Option 1 — VS Code Extension (recommended)

# Search in the Extensions panel
Extensions → Search "vexp" → Install

The extension auto-detects your AI agent and configures MCP automatically.

Option 2 — Standalone CLI

# Install globally via npm
npm install -g vexp-cli

# Or run the daemon directly
vexp-core mcp --workspace .

Works with Claude Code, Codex, Opencode, and more.

Pricing

Vexp offers three tiers:

| Plan | Price | Details | |---|---|---| | Starter | Free | Core features, small projects | | Pro | $19/month | Full features, larger projects | | Team | $29/month | Per user, for teams |

Deeper Analysis

Vexp’s approach is architecturally sound. The core insight — that AST-based dependency graphs produce far more relevant context than naive file-level context — is well-established in static analysis tooling. The “pivot node + token-budgeted expansion” model is a sensible heuristic.

The token reduction claim of 65-70% is plausible given the mechanism. A typical medium TypeScript project query might consume 18k tokens naively; a dependency subgraph targeting just the relevant call chain could plausibly hit 2.4k-5k tokens.

The local-only constraint is the strongest Differentiator. Unlike cloud-based context engines (Retool, context7), Vexp guarantees code never leaves the machine. For enterprise or privacy-sensitive projects, this is a decisive factor.

Practical Evaluation Checklist

• Local-first, no network calls — code stays on your machine
• AST-based dependency graph — not just keyword search
• Token-budgeted context — prevents context window overflow
• MCP-compatible — works with major AI coding agents
• No account required — install and use immediately
• VS Code first-class — official Marketplace extension

Security Notes

• All processing is local. No telemetry, no cloud backend required for core functionality.
• The Starter plan is free with no account — no email, no credit card.
• Code never leaves your machine unless you explicitly configure cloud features.

FAQ

Q: Does it work with non-TypeScript languages? A: The website focuses on TypeScript/JavaScript. The core AST parsing approach could theoretically extend to other languages, but TypeScript is the primary use case documented.

Q: How does it compare to Retool or context7? A: Retool and context7 are cloud-based. Vexp is entirely local. If data privacy is a concern, Vexp is the clear choice. If you prefer managed infrastructure, look elsewhere.

Q: Does it support Windows? A: Yes — the docs list Windows, macOS, and Linux as supported platforms.

Q: Is there a self-hosted option? A: The core engine runs locally. There is no separate self-hosted server to deploy — it runs as a CLI daemon or VS Code extension on your machine.

Conclusion

Vexp addresses a real pain point in AI-assisted coding: agents burning through context tokens on irrelevant code. By building an AST dependency graph and serving a targeted subgraph, it cuts token usage dramatically while keeping all processing local. If you use Claude Code, Cursor, or similar agents on non-trivial projects, Vexp is worth trying — especially given the free Starter tier and zero data leaving your machine.

Install it from the VS Code Marketplace or via npm install -g vexp-cli.