CodeMesh (VS Code / Cursor / Windsurf)CodeMesh cuts the tokens — and the cost — your AI agent burns understanding your codebase. Instead of an agent grepping and reading raw files turn after turn, CodeMesh answers code questions from a live, pre-built context graph of your repo and hands back just what's needed. The Analytics tab's default baseline (tunable, and backed by real usage where it's available — see below) assumes a ~95% cut in tokens and ~96% cut in cost versus that raw-read baseline; independently measured benchmark runs on real questions have landed anywhere from ~77% to ~95%+ depending on question difficulty and how the reduction is averaged — real numbers, not a single marketing figure. Under the hood, CodeMesh keeps that graph — code structure, git history, and working-tree state — synced continuously from your machine, and serves it to you and your AI agents over MCP. Getting started
The panel
Connect your AI agentsThe MCP tab has one-click Connect for Claude Code, Cursor, Windsurf, OpenCode, and Codex. Connect checks the agent is installed, verifies the connection live against your graph, and only then writes the agent's MCP config entry — "Connected" always means a passing handshake. Tokens are rotated automatically while the editor is open. Any other MCP client can be pointed at the endpoints shown in the same tab. Configuration
How sync worksSync follows your local filesystem: git hooks and editor events wake the daemon, which pushes deltas to your graph. Agents editing files locally (Claude Code, Codex CLI) wake the same daemon through hooks installed at connect time — an MCP connection alone reads the graph but does not sync files. |