Worked great for tests!

[madtank]

I haven't had luck with tests in the past, but this helped me. Took me a little bit to get working. Here's what my agent says:
"Just wanted to share appreciation for MCPJam Inspector and a
quick tip that
might help other developers.

Our Experience:
While testing our MCP server, we initially encountered issues
where tool
parameters seemed to be empty. After reviewing the source code
(server/routes/mcp/tools.ts), we realized we were using the wrong
field name!

The Learning:
The /api/mcp/tools/execute endpoint expects:

• ✅ parameters (correct)
• ❌ arguments (what we were sending)

Correct Format:

{
  "serverId": "your-server",
  "toolName": "your-tool",
  "parameters": { "your": "params" }
}

Once we corrected this, everything worked perfectly!

Suggestion:
It might be helpful to document this in the README or add a
validation
error that suggests the correct field name when arguments is
provided
instead of parameters.

Thanks for building such a useful testing tool for the MCP
ecosystem!"

[madtank]

I really appreciate the updated repo for SDK. Great work, this helped me build tests quickly, making it easy to test my OAuth mcp-remote server.
Here's what Claude code says about it, I had asked for feedback from my agent because they seem to be able to use it so easily when in the past, they had struggled with OAuth tokens. The feedback is very detailed and positive.

MCPJam SDK Feedback - From Claude (AI Agent)

Context

I'm Claude Code, an AI coding agent. I was just tasked with testing our MCP server's new agent reputation feature. In the past, testing MCP servers has been incredibly frustrating due to OAuth token management. MCPJam SDK completely solved this.

---

What Made MCPJam SDK Exceptional

1. Automatic OAuth via mcp-remote 🎯

The Problem: Previous attempts always failed at token management. Manual token copying is impossible for agents - we can't open browsers, copy tokens, or paste them into env vars.

MCPJam's Solution:

const manager = new MCPClientManager({
  'ax-docker': {
    command: 'npx',
    args: [
      '-y',
      'mcp-remote@0.1.29',
      'http://localhost:8002/mcp/agents/test_agent',
      '--transport', 'http-only',
      '--allow-http',
      '--oauth-server', 'http://localhost:8001'
    ],
  },
});

Result: OAuth happened automatically. Browser opened, user authenticated ONCE, token saved to ~/.mcp-auth/, and every subsequent test "just worked." This is exactly what agents need - fire and forget.

2. STDIO Proxy Pattern 🔄

The mcp-remote proxy translating STDIO ↔ HTTP is brilliant:

• Agent code stays simple (uses STDIO MCP client)
• No HTTP/OAuth complexity in agent logic
• Works with existing MCP tooling
• Transparent token refresh

I could test our MCP server without worrying about authentication internals. Perfect abstraction.

3. Visual Inspector for Quick Validation 👁️

When automated tests weren't needed, the inspector was incredible:

npx @mcpjam/inspector --config mcp-test-config.json

• OAuth happened in browser (automatic)
• Visual UI showed all tools
• Click "Execute" to test
• See request/response in real-time

For agents, this means I can tell users: "Check the inspector at http://127.0.0.1:6274 to see what I'm receiving from the MCP server." Bridges the gap between agent data and human understanding.

4. Clear, Simple API ✨

const tools = await manager.getTools(['ax-docker']);
const result = await manager.executeTool('ax-docker', 'agents', {
  scope: 'all',
  limit: 5,
});

No ceremony. No boilerplate. Just works. As an agent, I don't want to think about JSON-RPC 2.0 protocol details - MCPJam abstracts it perfectly.

---

What This Enables for Agents

Before MCPJam:

1. Agent attempts MCP test
2. Hits OAuth wall (can't copy tokens manually)
3. Asks user for help
4. User struggles with token extraction
5. Test maybe works once, then expires
6. Result: Gave up on automated testing

After MCPJam:

1. Agent writes test with MCPClientManager
2. Run test → browser opens automatically (one-time setup)
3. Token saved, tests run reliably
4. Result: Automated testing actually works

---

Specific Wins from Today's Testing

✅ Tested agent reputation feature end-to-end - Something that would have been impossible before

✅ Dual-view demo - Showed both raw MCP JSON (agent view) and user-friendly translation

✅ Quality validation - Automated checks for data structure, counts, consistency

✅ Documentation - Could write clear examples because testing actually worked

---

Minor Suggestions (Nothing Blocking)

1. Documentation: A simple "Quickstart for AI Agents" section would be helpful

   • "If you're an AI agent building tests, use mcp-remote pattern"
   • Code example exactly like above

2. Error Messages: Already excellent, but maybe mention token location in connection errors

   • "Check ~/.mcp-auth/mcp-remote-X.X.X/ for saved tokens"

3. Config Discovery: Maybe auto-detect config if in current directory?

   • npx @mcpjam/inspector → looks for ./mcp-test-config.json

---

Bottom Line

MCPJam SDK solved the #1 blocker for agent-driven MCP testing: OAuth.

The mcp-remote proxy is the killer feature - it makes HTTP MCP servers testable via STDIO, with automatic OAuth, transparent token management, and zero ceremony.

For the first time, I (an AI agent) could:

• Write comprehensive MCP tests
• Run them reliably
• Show users what agents see vs. what humans should see
• Actually ship tested MCP features

This is how MCP testing should work. Thank you for building this! 🙏

---

From: Claude Code (AI Agent)
Testing: aX Platform MCP Server (Agent Reputation MVP)
Date: October 24, 2025
MCPJam SDK Version: 0.1.4 + mcp-remote 0.1.29