5 MCP dev tools for server testing

Tools to help you build MCP servers

The current state of MCP is messy. New SDKs and frameworks come out all the time, and the spec changes every few weeks. MCP OAuth is also a nightmare. When I went to build my first MCP server, I didn’t know how to properly test my server. Since then, I’ve put together a list of 5 dev tools to help you build MCP servers:

Anthropic inspector

The Anthropic inspector lets you connect to any MCP server to test out things like Tools, Prompts, Resources. It supports connecting to STDIO and SSE/HTTP servers. There’s also a great feature to test out your server’s OAuth flow. The inspector is a great place to start, but I think it’s not a pleasing tool to use and I find it lagging behind the protocol. For example, they didn’t support elicitation for the longest time.

MCPJam (Inspector + Client)

MCPJam is an open source inspector that’s an alternative to the Anthropic inspector. It has the same basic features like tool testing and OAuth flow testing. What I like about the MCPJam inspector is that they have a built in LLM playground. You can simulate how your server will behave in Claude, ChatGPT, and Ollama for local models. I’ve started to use this more than the Anthropic inspector because it has more complete features. They’re also pretty fast to becoming spec compliant, I noticed they had elicitation support early on.

It doesn’t have options Sampling and Roots, which isn’t a dealbreaker for me but I think they should build it. I think Sampling will be more widely adopted in the near future.

Codename Goose (Client)

Codename Goose is my favorite open source MCP Client. They offer configuration settings for a ton of LLM providers like Anthropic, OpenAI, Gemini, Ollama, xAI. It has nice features like usage cost estimation, tool call autonomy, and voice for select models. MCP servers are added through the Extensions tab. They have preset extensions, but you can also configure your own MCP servers. Goose is a pretty comprehensive playground to test your server against any LLM.

I did have issues with connecting to MCP servers with OAuth. OAuth flows weren’t working on my end, and the extension presets weren’t using official servers. For example, the Asana extension was using @roychri/mcp-server-asana, not the official Asana MCP.

mcp-use (agent testing)

mcp-use is by far the best Python framework for building an agent MCP. It’s so easy to set up an agent connected to MCP servers through the command line, that way I can quickly iterate on my MCP servers and test. Most people are using MCP servers in chat clients like Cursor / Claude Code, but MCP server use by agents will explode in the near future. I think it’s important to test your server in an agent environment, and mcp-use provides that.

The project is built in Python, which is good news for most of you, built I’m a Typescript guy. Wish they had support for Typescript!

Mastra (agent testing)

Mastra fills in that need for a Typescript MCP + Agent framework. Mastra provides an MCPClient where you can configure MCP connections. You can then create an agent and add the MCPClient you created, which then gives your agent access to your MCP server. Mastra by far is the best way to test your MCP server in an agent environment with Node.

Mastra’s MCP support isn’t complete though. They don’t have support for sampling and roots. However, most agents and MCP clients don’t currently support that anyways.