Skip to content

Conversation

WongLoki
Copy link

Add SSE Transport Support for MCP Client

🎯 Overview

This PR introduces Server-Sent Events (SSE) transport support for the Model Context Protocol (MCP) client, expanding beyond the existing STDIO transport to enable HTTP-based MCP server communication.

What's New

Core Features

  • SSE Transport Implementation: New SseTransport struct in crates/cli/src/mcp_client/transport/sse.rs
  • Transport Type Selection: Enhanced CLI with --transport flag supporting stdio and sse options
  • Flexible Configuration: Support for both command-based (STDIO) and URL-based (SSE) MCP servers
  • Unified Client Interface: Seamless integration with existing MCP client architecture

Technical Changes

  • Dependencies: Added reqwest-eventsource and tokio-stream for SSE handling
  • CLI Enhancement: Updated mcp add command to support transport-specific parameters
  • Configuration Schema: Extended CustomToolConfig to handle multiple transport types
  • Client Architecture: Modified CustomToolClient enum to support both STDIO and SSE variants

🧪 Testing Status

⚠️ Important Notice: This implementation has been tested only in local development environment with the following verification:

Local Testing Completed

  • SSE connection establishment with deepwiki MCP server (https://mcp.deepwiki.com/sse)
  • JSON-RPC protocol compatibility verification
  • MCP initialization handshake validation
  • Basic transport layer functionality

🔍 Recommended Testing Before Merge

We strongly recommend the upstream team conduct comprehensive testing including:

  • Cross-platform compatibility (macOS, Linux, Windows)
  • Various SSE server implementations
  • Error handling and reconnection scenarios
  • Performance impact assessment
  • Integration testing with existing MCP workflows
  • Edge cases and network failure resilience

📋 Usage Example

# Add SSE-based MCP server
./cli mcp add --name deepwiki --transport sse --url https://mcp.deepwiki.com/sse

# Traditional STDIO server (unchanged)
./cli mcp add --name local --transport stdio --command "python server.py"

🔧 Implementation Details

Key Files Modified

  • crates/cli/src/mcp_client/transport/sse.rs - New SSE transport implementation
  • crates/cli/src/mcp_client/transport/mod.rs - Transport type definitions
  • crates/cli/src/cli/chat/tools/custom_tool.rs - Client configuration updates
  • crates/cli/src/cli/chat/mcp.rs - CLI command enhancements

Architecture

  • Maintains backward compatibility with existing STDIO transport
  • Uses reqwest-eventsource for robust SSE handling
  • Implements MCP protocol over HTTP POST + SSE event streams
  • Follows existing error handling and logging patterns

⚠️ Deployment Recommendation

Please treat this as a DRAFT implementation requiring thorough validation.

While functional in local testing, we recommend:

  1. Code Review: Thorough architectural review
  2. Extended Testing: Multi-environment validation
  3. Gradual Rollout: Consider feature flag or experimental status
  4. Documentation: User-facing documentation updates

🤝 Feedback Welcome

This implementation aims to expand MCP capabilities while maintaining system stability. We welcome:

  • Architecture feedback
  • Additional test case suggestions
  • Security considerations
  • Performance optimization ideas

Author: @WongLoki
Status: Ready for Review & Testing
Priority: Non-Breaking Enhancement

kensave and others added 30 commits May 16, 2025 18:26
* [feat]: Add a new semantic_search_client crate that provides vector embedding and
    semantic search capabilities for the Amazon Q CLI. This implementation:

    - Supports text embedding generation using Candle and ONNX runtimes
    - Provides hardware acceleration via Metal on macOS
    - Implements efficient vector indexing for semantic search
    - Includes file processing utilities for various file types
    - Supports persistent storage of semantic contexts
    - Includes comprehensive test coverage

    This crate will enable memory bank functionality for Amazon Q, allowing
    users to create, manage, and search through semantic memory contexts.

    🤖 Assisted by [Amazon Q Developer](https://aws.amazon.com/q/developer)

* Update semantic_search_client dependencies in Cargo.toml

* Refactor embedder implementation for Linux platforms to use trait objects

This change modifies the semantic search client to use Box<dyn TextEmbedderTrait>
on Linux platforms instead of directly using CandleTextEmbedder. This provides
more flexibility and consistency with the implementation on macOS and Windows,
allowing for better extensibility and polymorphic behavior across all platforms.

* Update Cargo.lock file

* Remove redundant CandleTextEmbedder import for non-macOS/Windows platforms

* fix(semantic_search): Update conditional compilation flags for embedders

Update conditional compilation flags to match the new embedding model selection logic:
- Replace target_env="musl" conditions with target_os conditions
- Update TextEmbedder trait implementation to use macOS/Windows condition
- Ensure consistent conditions across all files

🤖 Assisted by [Amazon Q Developer](https://aws.amazon.com/q/developer)

---------

Co-authored-by: Kenneth Sanchez V <[email protected]>
Co-authored-by: Kenneth Sanchez V <[email protected]>
* fix Build

* fix: Removes flakey test

---------

Co-authored-by: Kenneth Sanchez V <[email protected]>
- Tools and prompts are sorted by MCP server alphabetically. Built-in tools always come come first
- Within each category (MCP), they are sorted by name alphabetically.
- Slightly adjust /tools footer.
… as default (aws#1888)

- Removed ONNX embedder implementation and fastembed dependency
- Updated conditional compilation to use Candle on all platforms except Linux ARM
- Fixed duplicate implementations in error.rs
- Updated benchmark tests to remove ONNX references

🤖 Assisted by [Amazon Q Developer](https://aws.amazon.com/q/developer)

Co-authored-by: Kenneth Sanchez V <[email protected]>
* chore: update amazon clients

* chore: update windows-core in lockfile

* chore: exclude semantic search crate
* Revert "chore: exclude semantic search crate"

This reverts commit 6fe1e3e.

* fix(semantic_search): Fix VectorIndex is_empty implementation

The VectorIndex.is_empty() method was incorrectly checking if ef_construction
was equal to 100 to determine if the index was empty. This approach was
unreliable and caused test failures. This change adds a proper counter to
track the number of elements in the index and updates is_empty() to check
this counter instead.

🤖 Assisted by [Amazon Q Developer](https://aws.amazon.com/q/developer)

---------

Co-authored-by: Kenneth Sanchez V <[email protected]>
* precomputes total number of servers in loading prior to spawning display task

* changes display task to use async task instead of spawn blocking

* waits on notify instead of display task for initial loading

* includes time taken in warning and error

* adds slash command to show mcp server load messages

* includes mcps that fail to start in /mcp

* surfaces command error for mcp servers in non-interactive mode

* only surfaces timeout warning in non-interactive mode with one or more server in the client list

* adds copy change for /mcp

* uses a hash set of server names to keep track of the number of initialized servers

* only shows mcp non-interactive msg in non-interactive mode
* chore: add logs for refreshing token

* fix: save device registration when launching chat
* remove autocomplete

* fix cargo

* remove typos config

* remove mise and fix deny

* fix tests
- Introduced `reqwest-eventsource` and `tokio-stream` dependencies.
- Updated `CustomToolConfig` to include optional `url` and `transport` fields.
- Enhanced MCP server management to handle both STDIO and SSE transports.
- Updated client initialization to support SSE transport with appropriate error handling.
- Modified command-line arguments to accommodate new transport options.
@WongLoki WongLoki requested a review from a team as a code owner May 29, 2025 11:43
@kshwetabh
Copy link

I'm a Q-Developer Pro user, and the missing feature is blocking Figma MCP integration. Can this be prioritized?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

10 participants