Vexp — Context Engine for AI Coding Agents

58% Less Cost. 90% Fewer Tool Calls. 100% On Your Machine.

SWE-bench Verified: 73% pass@1 at $0.67/task · Docs · Changelog

The Problem

AI coding agents waste more than half their budget before writing a single line of code. They read files, run searches, open dependencies — 23 operations on average just to orient themselves. The result: higher costs, slower answers, context windows flooded with irrelevant code. And when the session ends, everything is forgotten — tomorrow the same agent rediscovers the same codebase from scratch.

What Vexp Does

Vexp pre-indexes your codebase into a dependency graph and delivers only the code that matters — ranked by relevance, bounded to your token budget. One install, and every AI agent you use starts informed instead of exploring blind.

• 23 tool calls → 2. The agent stops guessing and starts coding.
• $0.78 → $0.33 per task. Measured on SWE-bench Verified.
• 117s → 74s per task. Faster because there's less to process.

No accounts. No API keys. No code leaving your machine.

Benchmark Results

Tested on SWE-bench Verified — 100 real-world bugs, 4 agents, same model (Claude Opus), same budget.

On 7–10 tasks, Vexp was the only agent to produce a passing patch.

Full methodology and reproduction steps →

Quick Start

1. Install the Vexp extension — from this page, or search "vexp" in your editor's Extensions view
2. Open any project folder — Vexp indexes your codebase automatically
3. Ask your AI agent anything — Vexp delivers the right context, automatically

No account. No CLI. No API key. Working in under 60 seconds.

Not on stock VS Code? Vexp runs in the AI-first forks too — Antigravity, Cursor, Windsurf, Kiro, VSCodium (their Extensions views are powered by Open VSX, where Vexp is published) and Trae via "Install from VSIX". Same extension, same features.

Auto-Configures 14 AI Agents

On activation, Vexp detects the AI agents you use and writes their configuration for them — tool instructions included. Zero manual setup.

Any other MCP-compatible agent can connect to the same server manually. Your agents don't compete with each other — they all share one context layer and one memory.

How It Works

Your AI agent asks: "How does authentication work?"

Without Vexp:
  → 23 tool calls, reads 14 files, 18,000 tokens consumed
  → Cost: $0.78 · Duration: 117s

With Vexp:
  → 1 call to run_pipeline
  → Returns: 3 key functions (full code) + 8 related signatures (skeletons)
  → 2,400 tokens · Cost: $0.33 · Duration: 74s

Same question. 87% fewer tokens. Better answer. Everything runs on your machine — no server, no cloud, no network calls.

Memory That Compounds

Your AI doesn't start from scratch every time. Vexp observes what your agent explores, decides and learns — and keeps it across sessions, shared by every agent you use.

• Auto-capture — each interaction becomes a compact observation, saved automatically
• Stale-knowledge alerts — when code changes, related observations are flagged [STALE]
• Structural change tracking — not just "file changed": which functions were added, removed, modified
• Pattern & anti-pattern detection — recurring changes become project guidelines; dead-end explorations get flagged
• Cross-session, cross-agent — pick up days later, from a different agent, without losing a thing

Smart by Default

• Semantic search — finds validateCredentials() when you ask about "authentication", even if the word never appears
• Intent detection in 15 languages — "fix bug" focuses error paths, "refactor" returns blast-radius analysis
• Change coupling — files that ship together are linked together
• Security filter — .env, credentials and SSH keys never enter the index; .gitignore is respected

Optional Local AI — +30% Savings

Enable vexp-devmind, a code-specialised model that runs entirely on your hardware, for an extra +30% token savings on top of the baseline: sharper ranking, higher-quality memory, smarter prompt preprocessing.

• 100% local, GPU-accelerated — Metal on Apple Silicon, CUDA on NVIDIA, CPU everywhere else. Auto-detected.
• One-click install — the sidebar prompts you once (~3.5 GB, one-time)

34 Languages

Plus Markdown docs, indexed structurally. Vue/Svelte/Astro components are first-class: script blocks parsed with full symbol extraction, on real file line numbers. SQL covers your migrations and schema — tables, views, functions, triggers.

In Your Editor

• Sidebar panel — engine status, index stats, token savings, per-repo cards
• CodeLens & hover — "N dependents across M files" on exported symbols
• Status bar — live indexing progress and usage counter
• Multi-repo workspaces — one unified context graph across repos; cross-repo dependencies detected and navigable. Ideal for monorepos and frontend+backend splits.

Security

• 100% local processing — zero network calls, zero cloud dependencies
• Binary verification — SHA-256 checksums verified on activation
• Secret filtering — sensitive files and patterns excluded from the index
• .gitignore respected — only your project's files, nothing else

Your code never touches a server. Not ours. Not anyone's.

Platforms

Native binaries for Linux x64/arm64, macOS Intel/Apple Silicon, Windows x64 — no runtime dependencies. GPU acceleration (CUDA/Metal) applies to the optional local AI; the core engine runs on any modern CPU.

Pricing

All plans run 100% locally — no per-query costs, no surprise bills. Pro pays for itself in days: ~$0.45 saved per task adds up fast. Full plan details →

CLI Available

No VS Code required — the same engine works from the terminal with Claude Code, Codex, or any MCP-compatible agent:

npm install -g vexp-cli && vexp setup

→ vexp-cli on npm

Benchmark · Documentation · Changelog · vexp.dev