[kosesena]

🐛 Bug-fix PR

Draft: opened so the work on #7090 is visible. Two possible directions were discussed in the issue; this one needs no new dependency and no change to the existing langchain/langgraph pins. Happy to adjust if you'd prefer the other.

📌 Summary

LLM Chat cannot start on main after the MCP SDK v2 upgrade (#6868): the llmchat extra still installs langchain-mcp-adapters, which has no release compatible with mcp>=2.0.0.

🔁 Reproduction Steps

Closes #7090. Install with --extra llmchat; uv.lock resolves langchain-mcp-adapters==0.3.1 next to mcp==2.2.0, its import of mcp.shared.context.RequestContext fails, _LLMCHAT_AVAILABLE becomes False and POST /llmchat/connect returns 500 with "dependencies are missing".

🐞 Root Cause

• langchain-mcp-adapters 0.3.1 has no upper bound on mcp and breaks at import time with v2; 0.3.2 pins mcp<2.0.0, so bumping the floor would make the extra unresolvable.
• Upstream closed v2 support (Support for upcoming Python MCP SDK v2.0.0 langchain-ai/langchain-mcp-adapters#578) as superseded by langchain.mcp. That module ships from langchain 1.4.0, which requires langchain-core>=1.6 and langgraph>=1.2.11, above the <=1.5.1 / <=1.2.0 bounds the llmchat extra allows, and the namespace is marked beta.
• The guarded import turned the ImportError into a "missing dependencies" message, hiding the cause.

💡 Fix Description

• MultiServerMCPClient is replaced by a small _MCPToolSource in mcp_client_chat_service.py. It takes the same {"default": connection} mapping, lists tools with mcp.client.Client (following next_cursor), and wraps each tool as a LangChain StructuredTool (input schema = the MCP tool's inputSchema) that calls it over a fresh session, like the adapter did. An error result raises ToolException.
• Transports: streamable_http uses the gateway's existing mcpgateway.utils.streamable_http_compat.streamable_http_client (with headers), sse uses mcp.client.sse.sse_client, stdio uses StdioServerParameters.
• langchain-mcp-adapters is removed from the llmchat extra; uv.lock only loses that package.
• When the llmchat imports fail, connect(), initialize() and reload_tools() now report the original ImportError.
• Non-text tool content (images, resources) is not passed through yet; the previous path returned text for the agent as well. I can extend this if needed.
• The docs pages that mention langchain-mcp-adapters describe connecting external agents to the gateway and are left unchanged.

📏 Reviewability

• This PR has one clear purpose
• The linked issue is not labeled triage
• Unrelated bugs or improvements are tracked in separate issues/PRs
• Tests are included with the code they validate
• If AI-assisted, I understand and can explain the generated changes

🧪 Verification

Check	Command	Status
Lint (changed source file)	ruff, black, isort, pylint (10.00/10)	✅
New tests	pytest tests/unit/mcpgateway/services/test_mcp_client_chat_service_tool_source.py (in-process MCP v2 server: list tools with schema, call a tool, error → ToolException, transport selection)	✅
Chat service + routers	pytest tests/unit/mcpgateway/services/test_mcp_client_chat_service*.py tests/unit/mcpgateway/routers	✅ 1792 passed, 1 failed (test_sso_router.py::test_handle_sso_callback_keycloak_oversized_id_token_skips_hint_cookie, fails the same way on main)
Doctests	pytest --doctest-modules mcpgateway/services/mcp_client_chat_service.py	✅

Existing tests that patched MultiServerMCPClient now patch _MCPToolSource; three assertions on the old "dependencies are missing" text were updated to the new message.

📐 MCP Compliance (if relevant)

• Matches current MCP spec (uses the MCP SDK v2 client)
• No breaking change to MCP clients

✅ Checklist

• Code formatted (black, isort)
• No secrets/credentials committed