Build a Complete Assessment in Claude Desktop: The MCP Walkthrough

The interesting part of MCP is not that Claude can call a tool. It is that a tool chain with domain rules built in can hand you a finished deliverable—an assessment app with questions, scoring dimensions, a report, and visual styling—while you stay in the chat window.

This walkthrough does exactly that: connect FormLM to Claude Desktop, authenticate without leaving the conversation, generate a working assessment from one sentence, review what the agent built, and publish it. About ten minutes if your coffee is already brewing.

What you need before starting

• Node.js 18 or newer on the machine that runs Claude Desktop
• A FormLM account—free is fine; the CLI works with any plan
• An Access Token: sign in at formlm.me, open Account Settings, and copy the token (or use the email verification-code login later, entirely in chat)

npm install -g @formlm/cli
# verify it is on PATH:
formlm-cli --version

Attach the server to Claude Desktop

Claude Desktop reads one JSON file. Add formlm-cli mcp as a stdio server, and include the env.PATH line—macOS launches desktop MCP servers without your terminal's PATH, so a Homebrew-installed Node would otherwise fail with command not found:

{
  "mcpServers": {
    "formlm": {
      "command": "formlm-cli",
      "args": ["mcp"],
      "env": { "PATH": "/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin" }
    }
  }
}

Fully quit and relaunch Claude Desktop. A new client should now appear in the tools panel with eight FormLM tools attached—three for auth, one generator, one executor, plus snapshot, skill, and exec. Setting it up in Cursor, Codex CLI, Windsurf, or Cline follows the same shape; the per-platform config locations live in the install guide on GitHub.

Authenticate from the chat

Type naturally: "log in to FormLM with this access token: …". Claude calls auth_login, the token is stored in the CLI's local config, and auth_status confirms the session. No browser round-trip. Prefer no token at all? Ask for email login instead and Claude will run auth_email_code and wait for the verification code you paste back.

Generate the whole app in one sentence

Now the part that makes this more than a demo. Describe the instrument you actually need—Claude calls formlm_generate, which runs FormLM's server-side agent pipeline, and roughly a minute later you have a complete app:

you › Create a workplace stress assessment with 10 questions
     across 3 dimensions (workload, recovery, control), 1–5 scale,
     with score interpretation and a report for each result level.

claude › Running formlm_generate … done. The app has:
  • form     10 questions, 1–5 Likert items, 3 dimensions
  • scales   weights per dimension + band thresholds
  • report   per-band interpretation, radar chart
  • connect  page styling for the fill-in flow
  App ID: 4Lvur… — want a snapshot walkthrough?

Why the output is well-formed: before writing any configuration, the agent reads the same domain-rule documents the platform's own generators use—exposed as MCP resources (formlm://skills/form, scale, connect, report, expert, share) and available through the formlm_skill tool. The P0 constraints are baked into every tool description, which is what keeps a conversational build from drifting into malformed scales.

Review what the agent built

Never publish on trust alone—MCP makes auditing cheap. Ask for "a snapshot of every module" and formlm_snapshot returns the aggregated state of form, scales, connect, report, and share in one call. Have Claude walk you through it: why this dimension set, what the thresholds mean, which questions carry the most weight. This is the step where your domain judgment earns its keep—the agent drafts, you decide.

Iterate without touching the editor

Follow-up prompts become edits. "Replace question 7 with a recovery-specific item", "tighten the high-stress band to 13 and above", "make the report lead with the radar chart"—Claude routes each change through formlm_execute and the module commands underneath (field, scale, report, connect), with formlm_exec available for anything on the command whitelist. The web editor remains there if you want it, but it is no longer on the critical path.

Publish, then point your pages at it

End with "publish it and give me all the links": share publish goes live, and the URL list includes the fill-in page, the editor, and the data-management view—and the published form's Data API endpoint comes with it. From there the storage patterns in the framework guide and the export workflows in the data-automation walkthrough both start with this same URL.