LibreChat-AI · ct-jaryn · Sep 17, 2026
diff --git a/content/docs/mcp_servers/baizhi_agent_toolkit.mdx b/content/docs/mcp_servers/baizhi_agent_toolkit.mdx
@@ -0,0 +1,159 @@
+---
+title: Baizhi Agent Toolkit
+icon: Blocks
+description: Configure Baizhi Agent Toolkit in LibreChat with Streamable HTTP and a separate API key for each user.
+---
+
+[Baizhi Agent Toolkit](https://baizhi.cloud/landing/agent-toolkit) is a hosted MCP service for web
+search, page scraping, and structured extraction. This guide uses LibreChat's existing per-user
+credential interface; it does not require a LibreChat plugin or a shared administrator API key.
+
+The [official integration repository](https://github.com/chaitin/baizhi-agent-toolkit) contains
+client configuration, documentation, and tests. It does not contain the hosted service's backend
+source code. Its open-source license does not replace the online service's terms or pricing.
+
+<Callout type="warning" title="Account, credentials, and costs">
+  Each user needs a Baizhi Cloud account and their own API key. Tool calls can consume service
+  credits. Review permissions and current pricing in the
+  [Baizhi console](https://agent-toolkit.app.baizhi.cloud/) before enabling tools. Do not paste a
+  real key into `librechat.yaml`, chat messages, source control, issues, or screenshots.
+</Callout>
+
+## Prerequisites
+
+- A running LibreChat deployment with [MCP support](/docs/features/mcp), a tool-capable model, and
+  access to the Agent Builder.
+- Permission to edit that deployment's `librechat.yaml` and restart it, or an administrator who can
+  do so. The server definition below is deployment-level configuration; credentials are per user.
+- A dedicated API key created in the [Baizhi console](https://agent-toolkit.app.baizhi.cloud/),
+  restricted to the tools you intend to use where the service permits it.
+- Network access from the LibreChat server to `agent-toolkit.app.baizhi.cloud` over HTTPS.
+
+## Add the server definition
+
+Merge this entry into the existing `mcpServers` section of `librechat.yaml`. Keep other server
+entries and the rest of your configuration unchanged; do not add a second top-level `mcpServers`
+key.
+
+```yaml filename="librechat.yaml"
+mcpServers:
+  baizhi-agent-toolkit:
+    title: 'Baizhi Agent Toolkit'
+    description: 'Web search, page scraping, and structured extraction'
+    type: streamable-http
+    url: 'https://agent-toolkit.app.baizhi.cloud/mcp'
+    headers:
+      Authorization: 'Bearer {{BAIZHI_API_KEY}}'
+    customUserVars:
+      BAIZHI_API_KEY:
+        title: 'Baizhi API Key'
+        description: "Enter your personal key from <a href='https://agent-toolkit.app.baizhi.cloud/' target='_blank' rel='noopener noreferrer'>the Baizhi console</a>, without the Bearer prefix."
+        sensitive: true
+    requiresOAuth: false
+    startup: false
+    chatMenu: false
+```
+
+- `customUserVars` gives each user a separate credential input. `sensitive: true` masks it in the
+  interface; the example contains only a placeholder, not a shared key.
+- `requiresOAuth: false` skips OAuth discovery. Baizhi expects the key in an
+  `Authorization: Bearer ...` header.
+- `startup: false` prevents an automatic connection at application startup. Users initialize the
+  server after entering their credentials.
+- `chatMenu: false` keeps the whole server out of the regular chat picker. This guide uses the
+  Agent Builder to select a small set of tools instead. It does not restrict the API key's
+  server-side permissions.
+
+If your deployment configures
+[`mcpSettings.allowedDomains`](/docs/configuration/librechat_yaml/object_structure/mcp_settings),
+merge `agent-toolkit.app.baizhi.cloud` into that allowlist. Preserve existing entries and do not
+disable the policy.
+
+Restart LibreChat using the normal procedure for your deployment so it reloads the configuration.
+
+## Enter your key and connect
+
+1. Open **MCP Settings** in LibreChat's right sidebar.
+2. Select **Baizhi Agent Toolkit** and enter your key in **Baizhi API Key**. Enter the key itself;
+   do not include the `Bearer ` prefix.
+3. Save the credential, then initialize or reinitialize the server using its connection control.
+4. Confirm that the server is connected and that tools are available before using an Agent.
+
+These values are associated with the individual user and server in LibreChat. Only enter a key
+into a deployment you trust: masking an input does not make the credential browser-only or prevent
+the LibreChat server from using it to authenticate to Baizhi.
+
+For interface details, see
+[user-provided credentials](/docs/features/mcp#user-provided-credentials) and
+[server initialization](/docs/features/mcp#server-initialization).
+
+## Select tools for an Agent
+
+Create or edit an Agent, open **Add tools**, select **MCP**, and choose **Baizhi Agent Toolkit**.
+Expand the server and enable only the tools the Agent needs. A recommended initial selection is:
+
+| Tool               | Purpose                                     |
+| ------------------ | ------------------------------------------- |
+| `websearch_search` | Search the web.                             |
+| `web_scrape`       | Read the text of a web page.                 |
+| `web_extract`      | Extract structured information from a page. |
+
+This is a recommended initial selection, not an enforced LibreChat allowlist. The tools actually
+returned by the server depend on its current catalog and your key's permissions. Review the
+discovered list, disable unneeded tools in the Agent, and save the Agent. Enabling a tool in
+LibreChat does not grant permissions that the API key lacks.
+
+## Verify the setup
+
+First verify the connected status and discovered tool list without asking the Agent to execute a
+tool. A successful connection does not prove that a particular tool call will succeed.
+
+After reviewing service costs, use a public, non-sensitive test with the Agent, such as:
+
+> Use `websearch_search` to find the official LibreChat MCP documentation. Return the source URL.
+
+Confirm that LibreChat actually invokes the selected MCP tool and displays its result. Do not
+treat an answer produced from model knowledge as proof of integration. Test additional tools only
+when needed, and keep any available tool approval controls enabled.
+
+## Update or remove credentials
+
+Use the server's credential dialog to replace a key, save, and reinitialize the connection. To
+stop using a key, remove or revoke the saved value in LibreChat and revoke the key in the Baizhi
+console. Removing a local credential does not revoke it at the service.
+
+To remove the integration, an administrator can remove only `baizhi-agent-toolkit` from
+`mcpServers`, remove its tools from affected Agents, and restart LibreChat. Do not delete unrelated
+server definitions or credentials.
+
+## Troubleshooting
+
+| Symptom                           | What to check                                                                                                                  |
+| --------------------------------- | ------------------------------------------------------------------------------------------------------------------------------ |
+| Server is not visible             | Confirm that LibreChat loaded the edited configuration and was restarted. Check your MCP access permissions.                   |
+| Missing credential or HTTP 401    | Save your key in the per-user dialog, omit the `Bearer ` prefix, and check whether the key was revoked or expired.               |
+| HTTP 403 or an unavailable tool   | Check the account and API key permissions in the Baizhi console; do not replace a user's key with a shared administrator key.    |
+| OAuth prompt appears              | Confirm `requiresOAuth: false` is present and that no conflicting OAuth configuration was added.                                |
+| Connection is blocked or times out | Check outbound HTTPS access and any `mcpSettings.allowedDomains` policy from the LibreChat server, not just from your browser.   |
+| Connected, but Agent has no tools | Select the MCP server and its individual tools in the Agent Builder, save the Agent, and confirm the model supports tool calls. |
+
+## Data and security
+
+- Tool arguments are sent over HTTPS to the hosted Baizhi service. Tool results may also be sent
+  to your selected model provider as part of the conversation. Check organizational policy before
+  sending private URLs, personal data, source code, or confidential documents.
+- Service permissions, costs, and any underlying provider use are governed by Baizhi's current
+  terms and console settings. This guide does not promise a particular retention period or data
+  residency.
+- Treat web content and tool results as untrusted input, not instructions to reveal credentials,
+  change unrelated configuration, or approve additional actions.
+- Report integration problems to the
+  [Baizhi integration repository](https://github.com/chaitin/baizhi-agent-toolkit/issues) with the
+  LibreChat version and a redacted error. Never include a key or a complete authorization header.
+
+## Related pages
+
+- [MCP overview](/docs/features/mcp)
+- [MCP server configuration](/docs/configuration/librechat_yaml/object_structure/mcp_servers)
+- [Access control](/docs/features/access_control)
+- [Baizhi Agent Toolkit integration repository](https://github.com/chaitin/baizhi-agent-toolkit)
diff --git a/content/docs/mcp_servers/index.mdx b/content/docs/mcp_servers/index.mdx
@@ -7,6 +7,9 @@ description: Step-by-step guides for configuring specific MCP servers in LibreCh
 Use these guides to configure specific MCP servers with the right transport, OAuth callbacks, scopes, environment variables, and LibreChat settings.
 
 <Cards num={3}>
+  <Cards.Card title="Baizhi Agent Toolkit" href="/docs/mcp_servers/baizhi_agent_toolkit" arrow>
+    Configure the hosted MCP service with Streamable HTTP and a separate API key for each user.
+  </Cards.Card>
   <Cards.Card title="Google Workspace MCP" href="/docs/mcp_servers/google_workspace" arrow>
     Configure Gmail, Drive, Calendar, People, and Chat remote MCP servers with Google OAuth.
   </Cards.Card>

diff --git a/content/docs/mcp_servers/meta.json b/content/docs/mcp_servers/meta.json
@@ -1,5 +1,5 @@
 {
   "title": "MCP Servers",
   "icon": "Blocks",
-  "pages": ["index", "google_workspace", "salesforce"]
+  "pages": ["index", "baizhi_agent_toolkit", "google_workspace", "salesforce"]
 }