MCP (Model Context Protocol) is an open standard for connecting AI models to external tools and data sources. Runtype supports both built-in MCP integrations and custom MCP servers.
MCP servers provide tools that AI models can call during flow execution. Unlike simple HTTP tools, MCP servers offer:
Save an MCP server once, use it across all flows.
Response:
mcp:<server>:*)To attach every tool a saved server exposes — current and future — include a single wildcard entry in toolIds:
The wildcard is expanded into concrete tool ids at execution time, so:
toolConfigs, perToolLimits, and usage tracking still operate on real, fully-qualified tool ids.mcp:mynotion:* + mcp:mynotion:create_page), the wildcard wins and the curated picks are skipped.Use passthrough when you trust the server end-to-end and want the agent to pick up new capabilities without configuration changes. Use the curated form (specific tool ids) when you want an explicit, audited subset.
Pass server configuration inline for dynamic use cases.
MCP servers support multiple authentication methods:
MCP tool IDs follow this pattern:
Examples:
mcp:context7:resolve-library-id - Context7 documentation servermcp:myslack:send_message - Your saved Slack servermcp:github:create_issue - Your saved GitHub serverFor runtime servers, tool IDs become:
The reserved * toolName is the passthrough wildcard — see Passthrough Mode above:
Saved servers can have environment-specific configurations:
Specify environment in dispatch:
MCP tools work alongside other tool types:
Credential Security
allowedTools when possibletimeout value (max 60000ms)api_key, ensure headerName is correct/mcp/servers/{name}/tools to verify discoveryallowedTools filter if settools/list endpoint