Skip to content
mcpfy
Start deploying

inspector

Debug MCP traffic from any browser

Fire tool calls at your deployed server, read raw JSON-RPC frames, browse resources and preview widgets, against real clients, with no local setup.

Open InspectorSelf-host it

Our open source tools are used by developers at top companies

NorthwindApertureCogentLumenBrightpathSundialMeridianIroncladBlue HarborKestrel

How it runs

See exactly what the model saw

  1. 1

    Attach

    Point the Inspector at any deployment.

  2. 2

    Invoke

    Call tools with real arguments.

  3. 3

    Inspect

    Read the raw protocol frames.

  4. 4

    Replay

    Re-run any call after a fix.

Live JSON-RPC

Every frame in and out, timestamped and expandable, so protocol bugs stop being guesswork.

Cloud Inspectorsession #4f2a

San Francisco, CA

72°F

Partly cloudy · H 74° L 58°

now72°
2pm74°
3pm73°
4pm70°
5pm66°
6pm63°

→ tools/call

{
  "name": "display_weather",
  "arguments": { "city": "San Francisco" }
}

← result20041ms

{ "content": [{ "type": "resource" }] }

GPT-5.2

18/18

Claude

18/18

Gemini

17/18

Widget preview

Render your interactive views exactly as the host client will, before a user ever sees them.

acme-mcp · zsh

➜ npx create-mcpfy-app

✓ tools/chart-sales.ts

✓ widgets/Chart.tsx

✓ mcpfy.config.ts

➜ mcpfy dev

ready on :4141 · 6 tools registered

tools/chart-sales.ts

export const chartSales = tool({
  name: "chart_sales",
  input: z.object({ period: z.string() }),
});

Session replay

Walk a real user's conversation end to end and re-run the exact call that failed.

Tool calls · 24h24,732

last 24h ↑ 12.4%

By country34 regions

Tracesp50 42ms

  • tools/call display_weatherclaude142ms
  • resources/listchatgpt18ms
  • tools/call search_docscursor842ms
  • prompts/listcopilot12ms
  • tools/call fetch_urlclaude503

p50

38ms

p95

112ms

errors

0.2%

In the box

The debugger MCP was missing

1

Tool testing

Invoke with arbitrary arguments.

2

Resource browsing

Inspect everything you expose.

3

Prompt testing

Exercise prompts in isolation.

4

Zero setup

Runs hosted, local or self-hosted.

Stop debugging MCP with print statements

Start deployingBook a call

Explore the rest of the platform

MCP Hosting

GitHub-native builds, preview URLs per branch, production endpoints with TLS and custom domains, and instant rollback, purpose-built for MCP traffic and the clients that connect to it.

Learn more

Cross-client testing

Your server behaves differently in ChatGPT than it does in Claude or Cursor.

Learn more

Publishing checks

Six categories of checks mapped to marketplace requirements, each failure with fix guidance and an autofix flow, plus live end-to-end runs inside the real clients.

Learn more

Public chat

Every deployment gets a shareable chat surface you can embed anywhere you already have an audience: no client install, no configuration.

Learn more

Analytics

Tool calls, sessions, error rates and latency percentiles broken down by client, client version, protocol version and country, with discovery traffic excluded by default.

Learn more

Documentation

Guides, API references and deployment how-tos.

Learn more

Everything on mcpfy Cloud shares one pipeline.