MCP servers

Use Model Context Protocol (MCP) servers to connect flows and agents to external tools and data sources. Runtype lists the tools that each server exposes and passes each tool’s input schema to the model. You can save a server for reuse or pass its configuration at runtime.

This page covers calling external MCP servers from Runtype. To expose your own Runtype product as an MCP server instead, see Setting up an MCP surface.

Choose a server type

Choose a server type based on how you supply its configuration:

Use saved MCP servers

Save a custom MCP server to reuse its tools across flows and agents.

Add a saved server

Choose a dashboard or API workflow:

Discover tools from a saved server

To verify a saved server and list its tools, send a POST request:

curl -X POST https://api.runtype.com/v1/mcp/servers/customer_support/test \
  -H "Authorization: Bearer YOUR_API_KEY"

The API returns a response similar to the following:

{
  "success": true,
  "toolsDiscovered": 2,
  "tools": [
    {
      "name": "search_articles",
      "description": "Search support articles"
    },
    {
      "name": "create_ticket",
      "description": "Create a support ticket"
    }
  ]
}

Replace YOUR_API_KEY with your Runtype API key.

A failed discovery still returns HTTP 200. Check the success field, not only the status code:

{
  "success": false,
  "error": "Connection refused",
  "tools": [],
  "toolsDiscovered": 0
}

The error value varies with the failure. When no saved server has that name, the API returns 404 with the message “Custom MCP server not found”.

Discover tools from a URL

To list a server’s tools before you save it, send the URL to the discover endpoint:

curl -X POST https://api.runtype.com/v1/mcp/discover \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "url": "https://example.com/mcp",
    "auth": { "type": "bearer", "token": "YOUR_MCP_TOKEN" },
    "timeout": 30000
  }'

Replace the placeholders with your values:

• YOUR_API_KEY: Your Runtype API key.
• YOUR_MCP_TOKEN: The token that the MCP server accepts.

The request body accepts url (required), auth, timeout, and serverName. When you pass serverName without auth, Runtype uses the credentials of a server with that name that you saved. It doesn’t use credentials that a teammate saved. The API returns a response similar to the following:

{
  "url": "https://example.com/mcp",
  "tools": [
    {
      "name": "search_articles",
      "description": "Search support articles",
      "parametersSchema": {
        "type": "object",
        "properties": { "query": { "type": "string" } }
      }
    }
  ],
  "total": 1
}

The Runtype MCP server offers the same discovery through its discover_mcp_server_tools tool. For more information, see Runtype MCP server.

Use saved tools in a flow

Add saved MCP tool IDs to the tools object in a prompt step. The following examples use the customer_support server:

Use the TypeScript SDK to dispatch a flow with saved MCP tools:

import { FlowBuilder, RuntypeClient } from '@runtypelabs/sdk'
const client = new RuntypeClient({
  apiKey: process.env.RUNTYPE_API_KEY,
})
const result = await new FlowBuilder()
  .createFlow({ name: 'Customer support agent' })
  .prompt({
    name: 'Support agent',
    model: 'claude-sonnet-5',
    userPrompt: 'Find the open support tickets for this customer',
    tools: {
      toolIds: ['mcp:customer_support:search_articles', 'mcp:customer_support:create_ticket'],
    },
  })
  .run(client, { streamResponse: true })

This example reads the API key from the RUNTYPE_API_KEY environment variable.

Use the Python SDK to dispatch the same flow:

import os
from runtype import RuntypeClient
client = RuntypeClient(api_key=os.environ["RUNTYPE_API_KEY"])
for event in client.dispatch({
    "flow": {
        "name": "Customer support agent",
        "steps": [{
            "type": "prompt",
            "config": {
                "model": "claude-sonnet-5",
                "userPrompt": "Find the open support tickets for this customer",
                "tools": {
                    "toolIds": [
                        "mcp:customer_support:search_articles",
                        "mcp:customer_support:create_ticket"
                    ]
                }
            }
        }]
    }
}):
    print(event)

This example reads the API key from the RUNTYPE_API_KEY environment variable.

Send the same flow with cURL:

curl https://api.runtype.com/v1/dispatch \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "flow": {
      "name": "Customer support agent",
      "steps": [{
        "type": "prompt",
        "config": {
          "model": "claude-sonnet-5",
          "userPrompt": "Find the open support tickets for this customer",
          "tools": {
            "toolIds": [
              "mcp:customer_support:search_articles",
              "mcp:customer_support:create_ticket"
            ]
          }
        }
      }]
    }
  }'

Replace YOUR_API_KEY with your Runtype API key.

Use passthrough mode

To attach every tool that a saved server exposes, add a wildcard entry to toolIds:

{
  "tools": {
    "toolIds": ["mcp:customer_support:*"]
  }
}

The wildcard expands into concrete tool IDs at execution time. This behavior means:

• New tools from the MCP server are available without editing the flow.
• Approval rules, toolConfigs, perToolLimits, and usage tracking use concrete tool IDs.
• When toolIds contains both the wildcard and a curated ID for one server, Runtype uses the wildcard and ignores the curated ID.

Choose passthrough mode when you trust every tool on the server and want to receive new tools without changing the flow. Choose curated tool IDs when you need to review each tool.

Use the TypeScript SDK to attach passthrough tools:

const result = await new FlowBuilder()
  .createFlow({ name: 'Customer support agent' })
  .prompt({
    name: 'Support agent',
    model: 'claude-sonnet-5',
    userPrompt: 'Help me manage support requests',
    tools: {
      toolIds: ['mcp:customer_support:*'],
    },
  })
  .run(client, { streamResponse: true })

Send the same passthrough configuration with cURL:

curl https://api.runtype.com/v1/dispatch \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "flow": {
      "name": "Customer support agent",
      "steps": [{
        "type": "prompt",
        "config": {
          "model": "claude-sonnet-5",
          "userPrompt": "Help me manage support requests",
          "tools": { "toolIds": ["mcp:customer_support:*"] }
        }
      }]
    }
  }'

Replace YOUR_API_KEY with your Runtype API key.

Share saved servers

Saved MCP servers in an organization are shared with its members. Members can list and use them, view metadata, test connections, and discover tools. Runtype encrypts saved credentials and does not return credential values in API responses.

Server owners and organization admins can perform these actions:

• Edit the server configuration, including the URL, timeout, transport, and allowed tools.
• Rotate or replace credentials.
• Refresh or reconnect the OAuth2 connection.
• Turn the server on or off.
• Delete the server.

Other members can view server details but cannot manage them. A management request from another member returns 403 with this message: “This MCP server was added by a teammate. Only the person who added it or an org admin can change it.”

Use the following fields in GET /v1/mcp/servers and GET /v1/mcp/servers/:name responses to control management actions in client code:

When an organization has duplicate server names, an admin’s by-name DELETE request returns 409 with candidate IDs. Pass the optional id query parameter, as in DELETE /v1/mcp/servers/:name?id=SERVER_ID, to identify the registration to remove.

Use runtime MCP servers

Pass a server configuration inline when its URL or credentials vary by request.

Choose runtime servers

Use runtime MCP servers for these cases:

• Pass per-user credentials in a multi-tenant application.
• Test a server before saving it.
• Configure a one-off integration.
• Set a server URL at request time.

Configure a runtime server

Pass the server configuration in the prompt step’s tools object:

const mcpServers = [
  {
    id: 'user_notion',
    name: 'User Notion',
    url: 'https://example.com/mcp',
    auth: {
      type: 'bearer',
      token: userCredentials.mcpToken,
    },
    timeout: 30000,
    transport: 'streamable_http',
    allowedTools: ['search_articles', 'create_ticket'],
  },
]
const result = await new FlowBuilder()
  .createFlow({ name: 'User support agent' })
  .prompt({
    name: 'Support agent',
    model: 'claude-sonnet-5',
    userPrompt: 'Help me manage my support requests',
    tools: {
      mcpServers,
    },
  })
  .run(client)

Pass the per-user token from your application in userCredentials.mcpToken.

Use the Python SDK to pass the same runtime configuration:

mcp_servers = [{
    "id": "user_notion",
    "name": "User Notion",
    "url": "https://example.com/mcp",
    "auth": {
        "type": "bearer",
        "token": user_credentials["mcp_token"]
    },
    "timeout": 30000,
    "allowedTools": ["search_articles", "create_ticket"]
}]
for event in client.dispatch({
    "flow": {
        "steps": [{
            "type": "prompt",
            "config": {
                "model": "claude-sonnet-5",
                "userPrompt": "Help me manage my support requests",
                "tools": {
                    "mcpServers": mcp_servers
                }
            }
        }]
    }
}):
    print(event)

Pass the per-user token from your application in user_credentials["mcp_token"].

Send the runtime configuration with cURL:

curl https://api.runtype.com/v1/dispatch \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "flow": {
      "steps": [{
        "type": "prompt",
        "config": {
          "model": "claude-sonnet-5",
          "userPrompt": "Help me manage my support requests",
          "tools": {
            "mcpServers": [{
              "id": "user_notion",
              "url": "https://example.com/mcp",
              "auth": {
                "type": "bearer",
                "token": "YOUR_MCP_TOKEN"
              },
              "timeout": 30000
            }]
          }
        }
      }]
    }
  }'

Replace the placeholders with your values:

• YOUR_API_KEY: Your Runtype API key.
• YOUR_MCP_TOKEN: The token that the MCP server accepts.

An Agent accepts the same entries in its tools.mcpServers list, including an Agent that you define with defineAgent and converge with ensure. For an example, see Connect inline MCP servers.

Authentication types

Choose one of the following authentication types for a custom MCP server:

Replace every credential placeholder in the table with a value that the server accepts. Replace the name inside each {{secret:NAME}} reference with a managed secret name from your account.

Every credential field in auth (token, username, password, and each headers value) accepts a {{secret:NAME}} reference, which Runtype resolves from your managed secrets when it calls the server. Put MCP credentials here, not in a saved MCP tool’s config, which rejects {{secret:NAME}} references.

Use the headers type when a server requires two or more HTTP headers. Each header value can contain a {{secret:NAME}} reference, which Runtype resolves from your managed secrets when it sends the request. The headers field is valid only with type: "headers" and must contain at least one valid HTTP header.

A saved server’s auth header values can also send the verified identity of the end user that an agent run serves. Use {{_identity.token}} for a short-lived JWT that Runtype signs for each connection, or {{_tenant.id}} and {{_endUser.id}} for the plain IDs:

{
  "type": "headers",
  "headers": {
    "Authorization": "Bearer {{_identity.token}}",
    "X-Tenant-Id": "{{_tenant.id}}"
  }
}

Runtype fills in these values when it lists the server’s tools and when it calls a tool. The agent must have a tenancy strategy, and the server URL must use https. When a run has no verified identity, Runtype doesn’t call the server. The token’s aud is the server’s canonical resource URI, for example https://mcp.example.com/mcp, not its host name. For the conditions and the token’s claims, see Send the assertion to a saved MCP server.

Machine-to-machine OAuth2

For an MCP server that accepts OAuth2 client credentials, register the connection through the API. Runtype resolves your managed secrets and requests access tokens itself, so the client secret and access token never reach the model or sandbox code.

In an organization, only an organization admin can register a client-credentials server or change its credentials, URL, or transport. An API key that registers a server with managed-secret references needs the SECRETS:READ scope in addition to TOOLS:WRITE.

To register the server, create managed secrets named TEAMBRIDGE_CLIENT_ID and TEAMBRIDGE_CLIENT_SECRET in Settings > Secrets, and then send the following request:

curl -X POST https://api.runtype.com/v1/mcp/servers \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "teambridge_context",
    "url": "https://mcp.example.com/mcp/",
    "auth": {
      "type": "oauth2",
      "grantType": "client_credentials",
      "tokenUrl": "https://tenant.example.com/oauth/token",
      "clientId": "{{secret:TEAMBRIDGE_CLIENT_ID}}",
      "clientSecret": "{{secret:TEAMBRIDGE_CLIENT_SECRET}}",
      "audience": "https://api.example.com/",
      "scope": "read:api"
    }
  }'

Replace YOUR_API_KEY with your Runtype API key. Replace the example MCP URL, token URL, audience, and scope with your provider’s values. The two secret references name the managed secrets you created. audience, resource, and scope are optional. audience and the RFC 8707 resource parameter are distinct, and Runtype sends each one separately. Your token endpoint must accept client_secret_post client authentication and return a bearer token.

Bind the returned server.id to an agent’s config.tools, or a flow step’s config.tools, without repeating the auth configuration:

{
  "mcpServers": [
    { "id": "SERVER_ID", "url": "https://mcp.example.com/mcp/" }
  ]
}

Replace SERVER_ID with the returned provkey_ ID. Keep auth absent. Runtype resolves the saved registration in the execution’s organization and environment; its saved endpoint and credentials take precedence over the URL in the binding. Add allowedTools to select named tools, or use toolIds: ["mcp:SERVER_ID:*"] instead of mcpServers to bind without a URL.

Runtype requests a token on first use and reuses it until shortly before it expires. It requests a new token after expiry, or once after the MCP server returns HTTP 401. Runtype doesn’t reuse a token that has no expires_in value. Your token endpoint can receive more than one token request for the same server. When a secret is missing or the token endpoint fails, Runtype stops the request instead of calling the MCP server without authentication.

Rotate the managed secrets without editing each bound agent. To change the token URL, audience, or scope, send the complete auth object to PATCH /v1/mcp/servers/teambridge_context. The registry encrypts auth at rest and does not return credentials in list, create, or update responses. Token endpoints must use public HTTPS URLs and cannot redirect credential requests.

Configure client-credentials auth on the saved server only. Inline mcpServers[].auth and an external tool’s auth field don’t accept OAuth2. Bind the saved server by ID instead.

OAuth 2.1 authentication

Use the dashboard OAuth2 flow when an MCP server requires OAuth instead of a static token.

Detect OAuth2 on URL entry

On the dashboard, enter the MCP server URL. Runtype probes the server for OAuth2 metadata through RFC 9728 protected-resource discovery.

When Runtype finds OAuth2 metadata, the dashboard shows Connect with OAuth2 instead of manual token fields. It displays the discovered issuer and supported scopes.

Authorize in the browser

Click Connect with OAuth2. The dashboard opens a popup for sign-in and consent. Runtype uses PKCE for the authorization-code exchange, so the browser does not handle a client secret.

After you approve access, the popup closes and the dashboard discovers tools for selection.

Keep the connection active

When an access token expires and the server issued a refresh token, Runtype refreshes the token automatically. You don’t need to reconnect on a schedule. When the authorization server rejects the refresh token, for example because access was revoked, runs that use the server fail with the MCP_REAUTH_REQUIRED code. To restore the connection, reconnect the server from Tools.

Use discovery diagnostics

When Runtype does not detect OAuth2, expand OAuth2 not detected. View discovery details. The panel lists each discovery phase, the URL that Runtype probed, the response status, and any error. Use these details to troubleshoot an MCP server that supports OAuth2 but is not detected.

Find OAuth2 setup

Use the OAuth2 setup in either of these locations:

• On the Tools page, when you create an MCP Server tool.
• In the agent tool-selection modal, when you add a server while choosing tools for an agent.

Configured OAuth2 servers appear in the agent tool-selection modal with their discovered tool counts.

Tool ID format

Use the following format for a saved custom server tool ID:

mcp:SERVER_NAME:TOOL_NAME

For example:

• mcp:customer_support:search_articles: A tool on the saved customer_support server.
• mcp:customer_support:create_ticket: Another tool on the same saved server.

Replace SERVER_NAME with the saved server name and TOOL_NAME with the MCP tool name. Runtype stores a saved server name in lowercase, with spaces replaced by underscores. A server that you name Customer Support has tool IDs such as mcp:customer_support:search_articles and the path /v1/mcp/servers/customer_support.

Runtime server tool IDs use the custom_ prefix:

mcp:custom_SERVER_ID:TOOL_NAME

Replace SERVER_ID with the runtime server’s id value and TOOL_NAME with the MCP tool name.

Use the * value as the passthrough wildcard for saved servers:

mcp:SERVER_NAME:*

Saved server configuration

The POST /v1/mcp/servers and PATCH /v1/mcp/servers/:name requests accept the following fields:

To turn off a saved server without deleting it, send a PATCH request:

curl -X PATCH https://api.runtype.com/v1/mcp/servers/customer_support \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "isActive": false }'

Replace YOUR_API_KEY with your Runtype API key.

A turned-off server doesn’t appear in the GET /v1/mcp/servers list. To turn it back on, send { "isActive": true } to the same path.

PATCH doesn’t accept the environment query parameter. When one name has registrations in more than one environment, a PATCH request can change any one of them. To change a specific registration, delete it with the environment query parameter and create it again.

Runtime server configuration

Configure runtime MCP servers with the following fields:

Use environment-specific servers

Assign an environment to a saved server when you need separate development and production configurations:

curl -X POST https://api.runtype.com/v1/mcp/servers \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "customer_support",
    "url": "https://example.com/mcp",
    "auth": { "type": "bearer", "token": "YOUR_DEVELOPMENT_TOKEN" },
    "environment": "development"
  }'
curl -X POST https://api.runtype.com/v1/mcp/servers \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "customer_support",
    "url": "https://example.com/mcp",
    "auth": { "type": "bearer", "token": "YOUR_PRODUCTION_TOKEN" },
    "environment": "production"
  }'

Replace the placeholders with your values:

• YOUR_API_KEY: Your Runtype API key.
• YOUR_DEVELOPMENT_TOKEN: The token for the development server.
• YOUR_PRODUCTION_TOKEN: The token for the production server.

Add the environment to the dispatch request to select a saved server configuration:

{
  "options": {
    "environment": "production"
  }
}

A saved server without an environment applies to every environment. When a dispatch sets an environment, Runtype uses the server registered for that environment and falls back to the server without one. When a dispatch omits options.environment, Runtype uses only servers that have no environment.

To get or delete one registration of a name, add the environment query parameter, as in GET /v1/mcp/servers/customer_support?environment=production. Without the parameter, the request targets the registration that has no environment. If you create a second server with the same name and environment, the API returns 409 with a message such as “A custom MCP server named “customer_support” already exists for production environment”.

Limits

Runtype applies these limits to runtime MCP configuration:

A background run continues after the original request returns, such as an Agent Loop turn that runs in the background or a detached subagent run. A synchronous call holds the original request open while the tool runs, so Runtype caps its timeout at 30 seconds and logs the reduction.

Combine MCP tools with other tools

Add saved tools, built-in tools, runtime tools, and runtime MCP servers to one prompt. This tools object combines these tool sources:

const tools = {
  toolIds: ['mcp:customer_support:search_articles', 'builtin:openai_web_search', 'YOUR_TOOL_ID'],
  mcpServers: [
    {
      id: 'runtime_server',
      url: 'https://example.com/mcp',
    },
  ],
  runtimeTools: [
    {
      name: 'get_customer',
      description: 'Get customer information',
      toolType: 'external',
      parametersSchema: {
        type: 'object',
        properties: {},
      },
      config: {
        url: 'https://example.com/customer',
        method: 'GET',
      },
    },
  ],
}

Replace YOUR_TOOL_ID with the ID of a saved tool.

Protect MCP credentials

Use these practices to protect MCP credentials:

• Store saved-server credentials in the server configuration. Runtype encrypts them before storage.
• Keep runtime credentials out of logs and persistent storage. Use environment variables or {{secret:NAME}} references instead of hardcoded tokens.
• Set allowedTools to limit the tools that a server exposes.

Troubleshoot MCP servers

A saved server is required when a step references its tools in toolIds. When a required server can’t load its tools, the run stops with a message that names the server and a recovery action. Timeouts and temporary outages get at most one discovery retry. Runtype stops the run without trying model fallbacks, prompt fallbacks (including fixed-message fallbacks), or additional agent turns. Reconnect or repair the server before you run the flow or agent again.

Structured run errors include error.details.mcpFailure with the serverId, serverName, stage, code, retryable, and diagnosticId fields. Keep the diagnosticId value when you investigate a failed run. Browser voice tests display the same message and a Check SERVER_NAME connection button that opens the server in Tools. These messages don’t include upstream response bodies or credentials.

The stage field is one of configuration, token_refresh, connection, or tool_discovery. The following table lists the code values:

Use these checks to troubleshoot common MCP server failures:

API reference

Use these endpoints to manage saved servers and discover tools. When you call an endpoint with an API key, the key needs the listed scope:

Next steps

Continue with these guides: