You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi @franciscojavierarceo ! I'd like to implement MCP support for Feast as requested. Here's my proposed approach:
Implementation Overview:
Add optional MCP server support using fastapi_mcp library
Configuration-driven via feature_store.yaml with mcp_enabled: false as default
Non-breaking changes with graceful degradation when fastapi_mcp is not installed
MCP Tools Exposed:
get_online_features - retrieve features for entities
list_feature_views - enumerate available feature views
list_feature_services - enumerate available feature services
get_feature_store_info - basic store metadata
Configuration Example:
feature_server:
type: mcpmcp_enabled: true # defaults to falsemcp_server_name: "feast-mcp-server"mcp_server_version: "1.0.0"
This follows Feast's existing patterns for feature server configuration and maintains full backward compatibility. Would you like me to proceed with creating a PR?
Feast has three remote server configurations (online, offline, and registry), each with its own implementation. The question is: what functions does the MCP server expose? will be different server for each implementation or single server or just implementing online feature server ?
Im not very familiar with the MCP server yet and haven't looked at the PR, so please ignore the question is not relevant
Good question @redhatHameed! I should have been more explicit to limit it to the online feature server, though some of that does allow for retrieval of data from the offline store.
@redhatHameed
"Will it be different servers or single server or just online feature server?"
Single server: MCP enhances the online feature server only
Not separate servers: It's the same FastAPI application with MCP capabilities added
Online-focused: Currently only applies to online feature serving, not offline or registry operations
So MCP gives you all the standard online feature server functionality PLUS the ability for AI agents to interact with it through the Model Context Protocol.
Activity
YassinNouh21 commentedon May 31, 2025
Hi @franciscojavierarceo ! I'd like to implement MCP support for Feast as requested. Here's my proposed approach:
Implementation Overview:
fastapi_mcplibraryfeature_store.yamlwithmcp_enabled: falseas defaultfastapi_mcpis not installedMCP Tools Exposed:
get_online_features- retrieve features for entitieslist_feature_views- enumerate available feature viewslist_feature_services- enumerate available feature servicesget_feature_store_info- basic store metadataConfiguration Example:
This follows Feast's existing patterns for feature server configuration and maintains full backward compatibility. Would you like me to proceed with creating a PR?
franciscojavierarceo commentedon May 31, 2025
That'd be awesome!
YassinNouh21 commentedon May 31, 2025
PR is ready can u take a look @franciscojavierarceo
redhatHameed commentedon May 31, 2025
@franciscojavierarceo @YassinNouh21
Feast has three remote server configurations (online, offline, and registry), each with its own implementation. The question is: what functions does the MCP server expose? will be different server for each implementation or single server or just implementing online feature server ?
Im not very familiar with the MCP server yet and haven't looked at the PR, so please ignore the question is not relevant
franciscojavierarceo commentedon May 31, 2025
Good question @redhatHameed! I should have been more explicit to limit it to the online feature server, though some of that does allow for retrieval of data from the offline store.
YassinNouh21 commentedon May 31, 2025
@redhatHameed
"Will it be different servers or single server or just online feature server?"
Single server: MCP enhances the online feature server only
Not separate servers: It's the same FastAPI application with MCP capabilities added
Online-focused: Currently only applies to online feature serving, not offline or registry operations
So MCP gives you all the standard online feature server functionality PLUS the ability for AI agents to interact with it through the Model Context Protocol.