Vexp

AI coding agents get expensive when they retrieve too much context and slow when they retrieve too little. Vexp sits between your editor and the agent, parses the repo into a dependency graph on your own machine, and hands the agent only the files and symbols the current task actually depends on — sized to whatever token budget you set.

What Vexp actually does

It replaces "paste the whole file and pray" context with a locally-built index of every function, class, and type in the codebase. That index answers agent queries in seconds, so you send tight, dependency-aware capsules instead of the wall of unrelated code most editors ship by default.

• Parse your codebase into a graph of functions, classes, and types using tree-sitter indexing across 30+ languages

• Return only the files and symbols the agent actually needs, bound to your token budget and never an empty capsule

• Trace call chains and import paths across files to see exactly what each piece of code depends on

• Run entirely on your machine with zero network calls, so your source never leaves your local environment

• Store all indexes in .vexp/index.db, a local SQLite file that stays inside the repo

• Filter secrets by default so sensitive credentials are never included in any context capsule

• Expose 11 specialized MCP tools for context search, impact analysis, and logic tracing

• Pick up exactly where the agent left off, and flag stale context automatically when code changes

The features that matter

• Local dependency graph — Tree-sitter indexing builds a searchable map of the repo in seconds and updates as you code.

• Off-cloud by default — No source ever crosses the network, and the optional vexp-devmind add-on runs an on-device model to compress prompts before they reach the cloud agent.

• One MCP config, twelve agents — A single MCP server plugs into Claude Code, Codex, Cursor, Windsurf, and eight more coding agents; Vexp writes the config for you.

• Fast setup — Install through the VS Code extension or standalone CLI, then run a single guided command to configure indexes, agents, and git hooks.

Who it's built for

Engineers who already spend real money on AI coding agents and can feel the token bill grow every week — indie hackers shipping side projects on Claude Code or Cursor, backend teams running Codex against monorepos, and consultants juggling half a dozen client repos on the same laptop. The security-conscious will care about the local-only posture: regulated codebases and NDA client work stay on the machine.

The deal

License Tier 1 starts at $49 for 10,000 nodes, 1 seat, and 1 repo per workspace. Tier 2 jumps to 50,000 nodes, 3 repos, and a shared workspace and git index; Tier 3 pushes to 100,000 nodes and 5 seats with priority paths and analytics; Tier 4 unlocks unlimited nodes and 10 seats. Pick the tier that matches your repo size, not the one that matches your ambition.