What you’ll build

Add Fluenta to Claude Desktop’s MCP config so any conversation can call two new tools: fluenta_idea_x-ray_sandbox for free instant previews and fluenta_idea_x-ray for the full paid run. Same server, same tools, same auth as the Cursor recipe — only the config file location changes.

Prerequisites

Claude Desktop with Remote MCP support (any 2026 release). Earlier versions only handled stdio servers; if your Claude can’t see the tools after restart, you probably need to update.
A Fluenta API key with the read_write scope. Create one if you don’t have it — the free tier ships with 2,000 credits, exactly one paid X-Ray.

Step 1 — Edit the config

Claude Desktop reads its MCP servers from a JSON file. The path depends on your OS:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json
Linux: ~/.config/Claude/claude_desktop_config.json

Open the file (create it if it doesn’t exist) and merge this snippet. If you already have other MCP servers configured, just add the fluenta entry inside the existing mcpServers object — don’t replace the whole file.

claude_desktop_config.json

{
  "mcpServers": {
    "fluenta": {
      "url": "https://dev.fluenta.space/backend/api/v1/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_FLUENTA_API_KEY"
      }
    }
  }
}

Don’t commit this file. Your Bearer token sits in plaintext. If you pair-program or share dotfiles, keep this in a private location and template the rest of your config.

Step 2 — Restart Claude

Fully quit Claude Desktop and reopen it. The MCP server registry only refreshes on app startup. After it comes back, open a new conversation and ask: “What MCP tools do you have access to?” Claude should list the two Fluenta tools.

Tools missing? Three usual culprits: (1) JSON syntax error in the config — run cat claude_desktop_config.json | jq to validate, (2) Claude Desktop is too old to support Remote MCP — update, (3) the Bearer token is wrong — check your keys.

Step 3 — Free reconnaissance

Start every idea evaluation with the sandbox tool. Costs nothing, runs in seconds, gives Claude a verdict it can cite back to you:

conversation: sandbox call

> User (in Claude Desktop):
  I'm thinking of building an AI chrome extension that auto-summarises every Notion page you open. Quick read on whether this has legs?

> Claude (calls fluenta_idea_x-ray_sandbox):
  arguments: {
    "title": "AI auto-summary chrome extension for Notion",
    "description": "Browser extension detects Notion page loads and shows a 3-bullet AI summary in a sidebar. Free tier 10 pages/day, paid $5/mo unlimited."
  }

  → type: "xray_sandbox_preview"
  → verdict: "promising"
  → demand_signal: "high — 22k monthly searches for 'notion ai summary'"
  → competition: "saturated — 4 funded competitors, 2 free open-source"
  → top_risk: "Notion's own AI ships natively for $10/mo — differentiation needs a clear angle (speed? export formats?)"

> Claude: The sandbox calls this "promising" with one heavy caveat: Notion's
  own AI ships natively. Demand is real (22k/mo searches), but the wedge has
  to be sharper than 'AI summary'. Want me to run the full X-Ray to dig in?

The sandbox preview returns four signals: a verdict bucket (prime / promising / experimental / weak), a demand signal, competitive density, and the top failure mode. Use it liberally; spend nothing.

Step 4 — Submit a paid run

When sandbox results suggest the idea is worth digging into, ask Claude to run the full paid tool. Be explicit about consenting to the credit cost — baking the confirmation step into the conversation avoids accidental spend:

conversation: paid submission

> User: Yes, run the full X-Ray. Use a fresh title and bias the description toward the export-format angle.

> Claude (calls fluenta_idea_x-ray):
  arguments: {
    "title": "AI Notion summary → export to Markdown / Linear / Gmail",
    "description": "Chrome extension that summarises any Notion page and exports the summary as ready-to-paste Markdown, a Linear issue, or a Gmail draft. Optimised for moving Notion content into other tools, not just reading it. $9/mo paid tier."
  }

  → type: "xray_submitted"
  → analysis_id: "an_01HXY4PQRSTUV..."
  → status: "queued"
  → credits_used: 2000

> Claude: Submitted. The full run takes ~20 minutes. I'll check back when
  you ask, or you can read it any time at:
  https://fluenta.space/app/x-ray/an_01HXY4PQRSTUV...

The paid tool returns immediately with type: xray_submitted and an analysis_id. The actual run takes ~20 minutes wall-clock; credits are debited at submission, not on completion.

Step 5 — Read the result

Same tool, called again with the analysis_id instead of a fresh idea. Returns either xray_submitted with a progress percentage if the run is still in flight, or xray_result with the full structured report once complete.

conversation: read result

> User (40 min later): How's the X-Ray going?

> Claude (calls fluenta_idea_x-ray with { "analysis_id": "an_01HXY4..." }):
  → type: "xray_result"
  → status: "complete"
  → lrs: 81
  → verdict: "prime"
  → demand: { score: 88, evidence: [...] }
  → competition: { score: 71, top_competitors: [...] }
  → monetization: { benchmarks: [...] }
  → risk_flags: [...]
  → recommendation: "launch — wedge holds, no incumbent owns the export angle"

> Claude: Big update from the sandbox preview. With the export-format
  wedge, the LRS climbs from 'promising' to 81 (prime launch). Three
  things to know:
  1. Demand: 88/100 — surprisingly strong for a niche.
  2. No incumbent on this exact angle. Notion AI doesn't export.
  3. Risk flag: Chrome Web Store reviews are hostile to AI extensions
     in early 2026. Plan for slow review cycles.

Make Claude proactive

Claude Desktop doesn’t support project-level rules the way Cursor does, but you can drop a standing instruction into your User Style or pin a system prompt at the start of recurring conversations:

You have access to Fluenta MCP tools. Whenever I describe a
product idea, call fluenta_idea_x-ray_sandbox FIRST (it's free
and instant). Surface the verdict and the top risk. Only run
the paid fluenta_idea_x-ray tool when I explicitly say "run a
full X-Ray" — confirm the 2,000-credit cost before submitting.

Troubleshooting

“I don’t see any tools.” Fully quit Claude Desktop and reopen. Hot-reload doesn’t work; tools register at app startup only.
401 Unauthorized. Bearer token typo, or the key was revoked. Regenerate at /app/settings/api-keys.
403 Forbidden on the paid tool. Your key has read scope only. Issue a read_write key.
402 Payment Required. Out of credits. Top up to keep going.
Run stuck at progress=0 for >5 min. Backend at peak load — expected occasionally. The run is queued, come back in 30 min.

What to do next

The same config works for every Claude Desktop conversation across every project — one paste, set for life.
Read the X-Ray lifecycle for what happens server-side during the 20-minute run.
Try the Python MCP client recipe to drive these same tools from your own scripts.

Use Fluenta tools from Claude Desktop.