pingfusi
MCP server + CLI that puts a real human in your coding agent's loop. It publishes work mid-task, a reviewer pins what's wrong and returns a verdict, and the agent iterates until approved.
A CLI tool that adds a human review loop to AI coding agents, letting them clone websites pixel perfectly or get pinned feedback on any draft.
Pingfusi is a command line tool that adds a real human reviewer into the loop when an AI coding agent is building or cloning a website. Instead of an agent guessing whether its output looks right, it publishes a draft and files a review round. A human then looks at the result the way a person actually would, leaves comments pinned to the exact elements that are wrong, and gives a verdict the agent has to act on. The agent fixes the issues, resubmits, and this repeats until a human approves the final version.
The main advertised use is cloning an existing website so it matches pixel for pixel, including animations. The README frames this as solving a common problem: an AI agent can usually get a clone about ninety percent right, then a person wastes a lot of time nudging it through prompts to fix the remaining layout, font, color, and animation mismatches. With pingfusi, numeric checks catch what can be measured automatically, and a human review round catches the more subjective parts, like whether something simply looks off.
Setup for Claude Code, Codex, or Cursor users is a single command that installs the needed skills and tooling, after which a user can just tell their agent something like "clone example.com, use pingfusi." There is also a plain command line workflow for scaffolding a clone target, capturing the live site, building the clone, hosting a draft, and requesting a review, plus a simpler mode for reviewing an existing draft you already built yourself.
Animation handling runs by default during the build, recording what the real page's animations do and replaying them in the clone using recorded timing and parameters, with warnings rather than failures if something cannot be matched. Reviews and cloned results are billed per completed result, similar to metered API usage. The project is released under the MIT License.
Where it fits
- Have an AI agent clone a website and get it approved by a real human reviewer
- Ask a human a quick advisory question from the command line
- Get pinned comments and a verdict on any draft your agent publishes