Manufact
Test your MCP server across real clients
Run the same task in ChatGPT, Claude, the Inspector, Claude Code, Codex CLI and Pi. Get an LLM judge report and fix prompts for server issues it finds.
Already have an MCP server? Try connecting it below
Our open source tools are used by developers at top companies
Test tool use and MCP Apps UI
Describe a task and what success looks like, then choose clients and models. All clients test tool use. Browser clients can also render your MCP app's UI; coding-agent CLIs run without it.
Tool-use testing
Coding agents call your tools and complete the task. No app UI is rendered.
Codex CLI
PiIllustrative tool run
> Create an issue and assign it to me.
Tool arguments and results captured for evaluation.
Tool use + MCP Apps UI
ChatGPT, Claude and the Inspector render your app. We capture a screenshot for the judge and a recording for you to replay.
Compare results across clients
Review scores, model cost, tokens and duration for each run. Browser clients do not report model cost or token counts.
Turn findings into fixes
The judge assesses task completion from the recorded evidence. Review server findings and copy a fix prompt for your coding assistant, then run the suite again after your change.
Tool-use testing
Every client tests whether an agent can complete your task using your tools. Claude Code, Codex CLI and Pi run without rendering MCP Apps UI.
MCP Apps UI testing
The judge reviews a captured screenshot alongside the conversation and tool results. Add UI criteria to your rubric to check that the app renders with readable content and no clipped elements. Replay the recording to investigate.
Evidence and fix prompts
Review task-completion scores and the evidence behind server findings. Copy a fix prompt and compare results after your change.
Run your first test suite
Test a server hosted on Manufact or connect an existing MCP URL. Our blog post walks through creating a suite, reading results and fixing a tool issue.
