> ## Documentation Index > Fetch the complete documentation index at: https://docs.mcpcore.io/llms.txt > Use this file to discover all available pages before exploring further. # Playground > Chat with Claude or an OpenAI model in the browser and watch it call the tools of your MCP server. ## What it is for The **Playground** tab on the server detail page lets you talk to a real model that is connected to your server. For every answer you see which tool the model picked, the parameters it sent and the result your tool returned. Use it to: * Check that your tool **names and descriptions** lead the model to the right tool * See which **parameters** the model fills in, and whether it understands the optional ones * Try a tool end to end after you publish it, without setting up an AI client * Compare how Claude and OpenAI models use the same tools [Run](https://docs.mcpcore.io/tools/testing.md) in the tool editor tests one tool with values you type. The Playground tests the whole server the way an AI client uses it: the model decides which tools to call, in what order and with which values. *** ## Use your own API key The Playground uses your own key from **Anthropic** or **OpenAI**, and the provider bills the usage to your account. * The key is sent straight from your browser to the provider. MCPCore never receives or stores it. * By default the key is forgotten when you leave the page. Tick **Remember the key in this browser tab** to keep it until the tab is closed. * Anthropic and OpenAI each have their own key field. Switching the provider starts a new chat. Get a key from the [Anthropic Console](https://console.anthropic.com) or the [OpenAI platform](https://platform.openai.com/api-keys). *** ## Chat with your tools **1. Open the Playground tab** On the server detail page, open the **Playground** tab. **2. Pick a provider and model** Choose **Anthropic** or **OpenAI**, then pick a model from the list. See [Models](#models). **3. Enter your API key** Paste the key into the key field. **4. Ask something** Type a question, or click one of the suggestions built from the descriptions of your first tools. The model answers and calls your tools when they help. **5. Inspect the tool calls** Every tool call appears as a card in the answer, for example **Called** `get_order`. Click it to see the **Parameters** the model sent and the **Result** your tool returned. The model remembers the earlier messages of the chat, so you can ask follow-up questions such as *"and the one before that?"*. Click **New chat** to start over. ### Models | Provider | Model | Good for | | - | - | - | | Anthropic | Claude Opus 5.5 (default) | The best balance of quality and cost for testing tools | | Anthropic | Claude Fable 5.1 | Anthropic's most capable model, for the hardest multi-step tasks. Costs more | | Anthropic | Claude Sonnet 5.5 | Fast everyday use at a lower price | | Anthropic | Claude Haiku 5.5 | The fastest and cheapest Claude model | | OpenAI | GPT-6.1 Sol (default) | Complex work, close to GPT-6 Astra at a lower cost | | OpenAI | GPT-6.1 Terra | Fast and efficient | | OpenAI | GPT-6 Astra | OpenAI's most capable model | | OpenAI | GPT-6 Luna | The most efficient GPT-6 model for simple tasks | | OpenAI | GPT-5.6 Sol | Previous generation flagship | | OpenAI | GPT-5.5 | Previous generation | Which models your key can use depends on your provider account. If a model is not available for your key, the provider returns an error and the chat shows it. > [!TIP] > > To check that your tool descriptions work for every AI client, try the same question with a smaller model such as Claude Haiku 5.5 or GPT-6 Luna. If a small model picks the right tool, larger models will too. ### Reading a tool call card | Icon | Meaning | | - | - | | Green check | The tool ran and returned a result | | Red alert | The tool returned an error. The card shows **Error** instead of **Result** | | Clock | The call has no result yet | A failed tool call is not the end of the answer: the model sees the error and usually explains it or tries again with other values. That tells you how well your error messages help an AI client. *** ## Which tools the model sees * Every **active** tool on the server, in its **live** version. Drafts are not used until you [publish them](https://docs.mcpcore.io/tools/versions.md). * Tools marked **Deletes or overwrites data** in their [behavior](https://docs.mcpcore.io/tools/create.md#behavior) are turned off, see below. * The model reads exactly what an AI client reads: tool name, description and parameter schema. *** ## Safety * Tool calls run against your live server, your APIs and your databases, just like calls from any other AI client. They appear in [Traffic Logs](https://docs.mcpcore.io/observability/traffic-logs.md) and [Analytics](https://docs.mcpcore.io/observability/analytics.md), and count as normal requests to your server. * Tools marked as **Deletes or overwrites data** are turned off by default. When the server has such tools, a switch above the chat reads **Allow N tools that delete or overwrite data**. Turn it on only on purpose. While it is on, the Playground warns *"These tools run for real against your API and data."* * Tools marked **Changes data** or with no behavior set are available. Set the behavior of every tool that writes data, so the Playground and AI clients can treat it carefully. *** ## Server types | Security mode | Playground | | - | - | | Public | Works | | Unlisted | Works. The secret URL is sent to the provider only to connect | | API Key | Works. The server's API key is sent to the provider so it can connect | | Bearer Token | Not supported. The Playground has no JWT to send, so the provider cannot connect. Test from an AI client | | OAuth 2.0 | Not supported yet. Test from a client that supports OAuth, such as Claude Desktop or Cursor | The provider connects to your server from its own cloud over the public internet. The server must be active and its endpoint reachable. *** ## Messages and errors | Message | What to do | | - | - | | The API key was rejected. Check that it is correct and active. | Paste a valid key for the selected provider | | The provider is rate limiting this key. Wait a moment and try again. | Wait, or raise the limits of your provider account | | The provider could not connect to this MCP server. Make sure the server is active. | Check that the server is active and its endpoint works | | The model declined to answer this request. | Rephrase the request | | The answer was cut off because it reached the length limit. | Ask for a shorter answer, or split the question | When a request fails, your message is put back into the input so you can send it again. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.