CodeMesh - Cut AI coding costs by 90%已上线
Product overview
CodeMesh is a context-optimisation layer that sits underneath AI coding assistants such as Claude Code, Cursor and Windsurf. Instead of letting an agent re-read whole files on every question, it turns a repository into a searchable graph and serves only the relevant lines.
Target and users
Developers and small teams whose AI coding bills and latency grow because agents repeatedly read the same files — people who want to keep their favourite coding tool while making it cheaper and faster.
Page structure
Per the product's own docs and launch coverage: an editor extension (also listed on Open VSX), a background daemon that builds and maintains the code graph, an MCP connection for agents, and a terminal CLI as an alternative to the editor.
Core loop
Open a repository and the daemon builds a graph of files, functions, classes and calls. When an agent asks a question, CodeMesh serves just the relevant lines over MCP. Saving a file reparses only that file, tracked by blake3 content manifests, so a new chat session never has to relearn the project from scratch.
Features and pricing
- Editor extension plus CLI: `npm install -g @codemesh-app/cli`, then `codemesh login`, `codemesh sync`, `codemesh connect` (Node 18.17 or newer; the docs warn the unscoped `codemesh` npm package is unrelated)
- One account and one graph shared between extension and CLI
- Product Hunt lists the launch at 0 USD; detailed paid tiers were not stated on the launch page, so none are claimed here
Feature details
The maker's own published benchmark — 98 questions on the open-source openclaw repo — reports 95.1% fewer tokens (25.3M down to 1.23M) and API spend falling from $23.67 to $1.65 (maker's self-reported benchmark, with every question listed, per GetToolRank's summary of the project README/docs). Treat the headline "over 90%" saving as the maker's claim, not an independent test.
Visual and technical
Hosted service with editor extension, CLI and MCP server; proprietary, freemium (per GetToolRank, checked 2026-11-14 from the project's own pages). Works with Claude Code, Cursor, Windsurf and other MCP-capable agents named on its pages.
Not included / boundaries
Not a coding agent itself — it does not write code, it only supplies context to the agent you already use. It is not an open-source GitHub project (no GitHub stars listed), and its privacy note, per the same summary, says tool-call logs with IP addresses are kept indefinitely for now, with deletion by email request.
1VR Curated · Source: Product Hunt @CodeMesh
机器翻译 ·