Paste your server URL, pick a model, and watch it call your tools in a real conversation — DeepSeek, Claude, GPT-5, Gemini, Grok, Qwen, GLM, Kimi, MiniMax, Nemotron or Mistral. MCP is model-agnostic, so all of them can call your tools; what differs is whether a given model picks the right one and gets the arguments right. That is the part you can only find out by running it.
Paste a server URL, pick a model, and watch it call your tools in a real conversation. No install.
No server of your own? Leave it blank and use one of the public mock servers.
In your browser
MCP Agent Studio
Server has a public URL? Chat with 90+ models on it, watch every tool call, and compare models side by side.
Open MCP Agent Studio →On your computer
MCP Playground Desktop
Server runs on localhost or over stdio? The browser can’t reach it. The desktop app can, with 90+ models in Agent Studio. Download free
| Model family | Vendor | MCP support | In Playground | Run it |
|---|---|---|---|---|
| Claude | Anthropic | Native — the API accepts MCP servers directly | 12 models | Test with Claude → |
| GPT | OpenAI | Via an MCP client (tool-calling translation) | 18 models | Test with GPT → |
| Gemini & Gemma | Via an MCP client (tool-calling translation) | 10 models | Test with Gemini & Gemma → | |
| Grok | xAI | Via an MCP client (tool-calling translation) | 7 models | Test with Grok → |
| DeepSeek | DeepSeek | Via an MCP client (tool-calling translation) | 5 models | Test with DeepSeek → |
| Qwen | Alibaba | Via an MCP client (tool-calling translation) | 12 models | Test with Qwen → |
| GLM | Z.AI | Via an MCP client (tool-calling translation) | 9 models | Test with GLM → |
| Kimi | Moonshot AI | Via an MCP client (tool-calling translation) | 5 models | Test with Kimi → |
| MiniMax | MiniMax | Via an MCP client (tool-calling translation) | 4 models | Test with MiniMax → |
| Nemotron | NVIDIA | Via an MCP client (tool-calling translation) | 3 models | Test with Nemotron → |
| Mistral | Mistral AI | Via an MCP client (tool-calling translation) | 1 models | Test with Mistral → |
Run one prompt across several models at once
Same MCP server, same prompt, models side by side — see which one calls the right tool. Free account, no card.
Anthropic authored MCP, and the Claude API is the only one that speaks it natively rather than through a tool-calling shim. Strongest default for long multi-tool chains.
Reaches MCP through the Agents SDK and ChatGPT connectors. Broadest ecosystem support; tool-selection behaviour differs noticeably from Claude on identical schemas.
Test your server with GPT →Testing MCP with ChatGPT & the OpenAI MCP tool
Large context windows make Gemini forgiving of big tool lists, which is exactly when tool-selection drift is hardest to spot without a side-by-side run.
Test your server with Gemini & Gemma →Testing MCP with Gemini models
Fast and cheap per call, including a multi-agent variant worth testing against parallel tool use.
The cheapest credible tool-caller here. Worth benchmarking before paying for a frontier model on a routine workflow.
Test your server with DeepSeek →Connecting your MCP server to DeepSeek V4
The widest spread of sizes on this list, from Flash up to Max — useful for finding the smallest model that still calls your tools correctly.
Test your server with Qwen →Testing MCP with Alibaba Qwen models
Several GLM builds are free on the router, which makes them a cheap first pass when you just need to know whether your tools are callable at all.
Strong on long agentic runs; the Code variant is the one to try against developer-tool MCP servers.
A recent entrant to agentic tool use — thin third-party benchmarking so far, which is a reason to test rather than trust a leaderboard.
Test your server with MiniMax →Connecting your MCP server to MiniMax M3
Open-weight models including free-on-router builds. The realistic option if you need to self-host the model behind your MCP server later.
European-hosted option; useful where data residency decides the model before capability does.
Only Claude accepts MCP servers directly at the API level. Every other provider reaches your tools through an MCP client that translates your tool definitions into that provider's own function-calling format. Two things fall out of that:
You can reproduce both on demand: the complex-schema mock server exposes four tools with overlapping argument shapes specifically to provoke tool-selection drift.