> ## Documentation Index
> Fetch the complete documentation index at: https://docs.getmcp.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Chat Playground
> Test your GetMCP servers with a real AI client — Claude, ChatGPT, Groq, or Gemini — directly inside the WordPress admin.
The **Chat Playground** is a fully-featured chat UI built into the GetMCP admin. Drop in an API key for any of the four supported providers, pick a model + an MCP server, and start asking questions. Tool calls fire against your real GetMCP server with your real auth, results render inline, and the assistant streams a final answer — exactly how Claude Desktop, Cursor, or any other MCP client behaves.
Use it to:
* **Validate a new server end-to-end** before you involve a desktop client
* **Compare provider behavior** — ask the same prompt to Claude, GPT-4o, Llama, and Gemini and see which one nails your tool surface first try
* **Demo MCP to stakeholders** without "let me share my screen with my desktop app" friction
* **Iterate on a tool description or schema** — edit in the admin, reload chat, watch the assistant react in seconds
The Chat Playground is admin-only. Site administrators have access by default; other users must be added via the allowlist (covered below). API keys are encrypted on disk and only ever decrypted for your authenticated admin session.
***
## Open the Playground
In the WordPress admin, navigate to **GetMCP → Chats**.
The first time you open it, you'll see two things to configure: an **AI provider API key** and **MCP server credentials** (only if your server is auth-protected).
***
## Configure an AI Provider
Click the **gear icon** in the chat topbar to open Settings, then switch to the **AI Providers** tab. Four providers are supported:
| Provider | Models | API key from |
| ---------------------- | ----------------------------------------- | -------------------------------------------------------------------- |
| **Claude** (Anthropic) | Claude Sonnet 4.5, Opus 4.1, Haiku 4.5 | [console.anthropic.com](https://console.anthropic.com/settings/keys) |
| **ChatGPT** (OpenAI) | GPT-4o, GPT-4o mini, GPT-4 Turbo, o1 mini | [platform.openai.com](https://platform.openai.com/api-keys) |
| **Groq** | Llama 3.3 70B, Llama 3.1 8B, Mixtral 8x7B | [console.groq.com](https://console.groq.com/keys) |
| **Gemini** (Google) | Gemini 2.5 Pro, 2.5 Flash | [aistudio.google.com](https://aistudio.google.com/app/apikey) |
Expand any provider, paste the key, and click out. The accordion shows a green **Connected** badge once a key is saved. Keys are encrypted at rest with libsodium.
Calls go **directly from your browser** to the provider — they never pass through the GetMCP backend. This saves a server hop on every turn and means your prompts and responses aren't logged anywhere on the WordPress side beyond what the MCP server itself records.
### Anthropic prompt caching
For Claude conversations the chat automatically attaches `cache_control: { type: 'ephemeral' }` to every request. After the first turn, the conversation prefix (tools + system + every prior message) is served from Anthropic's cache at **\~10% of normal input cost**. OpenAI's automatic prefix cache kicks in too once the prefix exceeds 1024 tokens. No configuration needed.
***
## Configure MCP Server Credentials
Switch to the **MCP Credentials** tab in Settings. Every connected MCP server gets its own accordion with an auth-method selector:
| Method | When to use |
| -------------------- | ---------------------------------------------------------------------------------------------- |
| **None** | Open / public test servers |
| **Bearer Token** | Static bearer tokens — paste once, the chat sends `Authorization: Bearer …` on every tool call |
| **API Key Header** | Custom header name + value (e.g. `X-API-KEY`, `X-FlowMattic-Key`) |
| **Basic Auth** | Username + password, base64-encoded into `Authorization` |
| **OAuth 2.1 (PKCE)** | Full discovery → DCR → PKCE → token flow. Same flow ChatGPT and Claude Desktop use |
For all methods except OAuth, the token / value is stored encrypted at rest. For OAuth, see the next section.
***
## OAuth Connect Flow
When you select **OAuth 2.1 (PKCE)** and click **Connect**, the chat performs the full MCP OAuth flow:
The chat sends an unauthenticated `tools/list` call and reads the `WWW-Authenticate: Bearer resource_metadata="…"` hint from the 401 response.
Fetches RFC 9728 protected-resource metadata, then RFC 8414 authorization-server metadata to learn the `authorization_endpoint`, `token_endpoint`, and `registration_endpoint`.
POSTs an RFC 7591 client registration with the chat's stable redirect URI: `{site}/wp-admin/admin.php?page=getmcp-chats&getmcp_oauth_cb=1`.
Generates a code verifier + S256 challenge, persists the flow in `sessionStorage` keyed by `state`, then navigates the browser to the authorize URL with `resource={server endpoint}`.
After the user grants consent, the AS redirects back to the chat URL with `?code=…&state=…`. The chat exchanges the code at the token endpoint and persists the encrypted token bundle to `getmcp_chat_settings`.
On any subsequent 401 from the MCP server, the chat silently calls the token endpoint with the refresh token, swaps the access token in place, and retries the original call — without bothering you.
You'll see a green **Connected** state on the credential card once tokens are stored. To start fresh, click **Disconnect** (clears the access token, keeps the registered client) or change the auth method dropdown to None and back (forces fresh DCR).
***
## Send Your First Message
Back on the main chat view:
1. Pick a provider + model from the **first dropdown** in the topbar.
2. Pick an MCP server from the **second dropdown** (shows the tool count as a chip).
3. Type a question in the composer.
4. Press Enter (or click the send button).
The assistant picks a tool based on your prompt, the chat fires a JSON-RPC `tools/call` against your MCP server with the configured auth, captures the response, and feeds it back. The whole round-trip renders inline:
* **Tool trace** — "Discovering tools → Selecting → Calling"
* **Tool card** (expandable) — input arguments + raw JSON output + latency
* **Final answer** — streamed character-by-character if streaming is enabled
***
## Settings Reference
The Settings modal has three tabs:
### General
Behavior toggles, all persisted server-side per-site.
| Setting | Default | What it does |
| --------------------------- | ------- | --------------------------------------------------------------------------- |
| Show suggestions | On | Server-aware prompt suggestions in the empty state |
| Suggestion count | 4 | How many suggestion cards to show (1–6) |
| Show tool trace | On | Streams the "discovering → selecting → calling" trace before each tool call |
| Auto-approve tool calls | On | When off, the chat pauses and asks before every tool call |
| Expand tool card by default | On | New tool cards open expanded; off keeps them collapsed |
| Stream replies | On | Typewriter-stream the final answer; off drops it in all at once |
| Show raw JSON output | On | Show the "View raw JSON" toggle on each tool result |
### AI Providers
Provider keys (covered above).
### MCP Credentials
Per-server auth (covered above).
***
## Allowlist: Who Can Chat?
Chat is gated to a per-site allowlist. Site administrators (anyone with `manage_options`) always have access — they can't lock themselves out. Other admin-tier users (editors, etc.) need to be granted access individually.
In the WordPress admin, navigate to **GetMCP → Settings**.
Find the **Chat playground access** section.
Check the box next to each user's name. Use the filter to find people fast on busy sites.
Click **Save Changes**. The Chats menu item appears in the React sidebar for allowlisted users; for everyone else the menu is hidden and the REST endpoints return 403.
The list lives in `getmcp_settings.chat_allowed_users` as an array of user IDs.
Chat sessions consume your AI provider's API quota and execute MCP tools using stored credentials. Only grant access to users you trust with billable AI calls and tool-execution rights.
***
## Audit Trail
The chat identifies itself to your MCP server via the standard MCP `initialize` handshake with `clientInfo.name: "GetMCP Chat"`. Every tool call shows up in your call logs alongside Claude Desktop, ChatGPT, Cursor, and other clients — same logs page, filter by client type, replay any historical call inline.
***
## Troubleshooting
Either you're not logged in as a site administrator and haven't been added to the allowlist, or the chat plugin file is in an inconsistent state. Site admins always have access — if you're one and the menu is missing, deactivate + reactivate the plugin to refresh the menu registration.
A tool's input schema is missing `type: "object"` at the root. The chat normalises this automatically — make sure you're on plugin version 1.0 or later. If you wrote the schema by hand in the tool editor, add `"type": "object"` and a `"properties": {}` entry.
The redirect URL must be **fragment-free** — the AS appends `?code=…&state=…` to the registered URI, and if that URI already has a `#fragment`, the query params get trapped inside the fragment and never reach the client. The chat handles this correctly by default. If you previously connected from an older build, click **Disconnect** then **Connect** again to force a fresh dynamic client registration with the corrected redirect URI.
Your MCP server is OAuth-protected, but the access token has expired and refresh failed (or the credential is misconfigured). Open Settings → MCP Credentials, click **Reconnect** on the server card to run the full OAuth flow again.
Your MCP server's `auth_type` is set to `oauth` — Bearer tokens won't validate against it. Either switch the server's `auth_type` to `bearer` in the admin, or use the Chat's OAuth flow.
The Chat REST endpoints require both `getmcp_manage_servers` capability AND allowlist membership. If you're logged in as a non-admin user who isn't on the allowlist, saves are silently rejected. Check the Network tab for 403 responses.
***
The Chat Playground is included in every GetMCP install — no add-on, no extra license. Bring your own AI provider key and you're one click from a real AI client running against your real MCP server.