vexp — AI Context Engine for Claude Code, Cursor & Copilot

Right context in. Verified work out. 100% on your machine.

Your AI agent wastes most of its budget exploring your codebase before writing a line of code.
vexp pre-indexes your project locally and feeds any agent only the relevant code, then checks the work when it's done. On SWE-bench Pro: Fable-5-level results from Opus 5, at half the token price.

Install and you're done

Install the extension, open your project. vexp detects your AI agents, writes their config, builds a local graph of your code, and stays out of your way. No login. No API key. No CLI to install. One click and every AI agent in your editor gains context awareness.

Works out of the box with Claude Code, Cursor, Windsurf, GitHub Copilot, OpenAI Codex, Zed, Cline, Aider, opencode and more — 14 agents in total.

What it does

📉 Fable-5-level results from Opus 5. At half the token price.

On the SWE-bench Pro public set, Claude Code running Claude Opus 5 with vexp resolved 81.7% of 731 real GitHub issues (597/731). Anthropic reports 80.3% for Claude Fable 5, whose tokens cost twice as much, and 79.2% for Opus 5. That is $1.52 per task, with a 95% confidence interval of 78.7-84.3 and every session published: see the full benchmark. Since 3.0 the orientation vexp injects at turn zero carries up to three times more evidence, so your agent starts editing sooner instead of exploring, and everything vexp does that could add a turn to your session is opt-in: by default, vexp only subtracts.

⚡ Feeds your agent only the code that matters

AI agents read whole files and drown their context window in code that doesn't matter. vexp builds a dependency graph of your project with tree-sitter and returns just the pivot files — the exact functions, classes and call sites the task touches, with line ranges and blast radius. The result: your agent stops rediscovering the codebase before every change. In the context-capsule demo, a single call answers with 2.1k tokens instead of 8.2k (-74% on that call).

🔍 Verifies the work when the agent says it's done (new in 2.5)

Agents love to declare victory while leaving a broken import three files away. vexp's mechanical check reports what's provably broken or incomplete — parse errors, imports broken by a rename, dependents that were never updated, files the task promised and never produced — each with file and line. No model grades anything; it's pure git + tree-sitter + code graph.

🛡️ Shows what your code exposes to AI (new in 2.5)

Run vexp: Shield Scan from the Command Palette. vexp reports the emails, API keys, tokens, IBANs and private keys sitting in your comments and string literals that every AI agent reading your code currently sees — deterministic, local, masked in the report.

🩺 Diagnoses itself

vexp: Doctor in the palette checks daemon health, version skew, license and hooks in one place. Updates surface a one-click restart when needed.

🔒 100% on your machine

Zero network calls. Zero cloud indexing. Your code never leaves your laptop — a requirement for fintech, healthcare, defense, or any team with IP worth protecting, not a nice-to-have.

Command Palette

• vexp: Shield Scan — find PII and secrets any AI agent can read
• vexp: Doctor — diagnose daemon, version, license, hooks
• vexp: Restart Daemon — clean restart after an update

Supported languages

TypeScript, JavaScript, Python, Rust, Go, Java, Kotlin, Scala, C, C++, C#, Ruby, PHP, Swift, Objective-C, Lua, Dart, Elixir, Clojure, Haskell, OCaml, F#, Zig, Bash, SQL, and more — 34 languages via tree-sitter.

Pricing

Free for a personal project. Pro and Team plans unlock larger graphs, multi-repo workspaces, and the savings dashboard. Everything runs on your machine on every plan.

→ vexp.dev  ·  Docs

_{Keywords: AI context engine · Claude Code context · Cursor context · Copilot · save tokens · reduce AI coding cost · lower AI session cost · AI token optimization · MCP server · code verification · PII scanner · dependency graph · tree-sitter · local-first · offline · privacy}