MCP server semble_search tool times out on first queries (CLI works fine)

[cycle2zhou]

Environment

• Semble version: v0.1.10
• OS: macOS (darwin)
• Agent: OpenCode (Claude Code-compatible MCP client)
• Python/uv: uvx via uvx --from "semble[mcp]" semble
• Repo size: Java project, ~48 relevant files for JDBC module

Description

When using the Semble MCP server through an agent harness (OpenCode), the semble_search tool consistently returns MCP error -32001: Request timed out on the first several queries. The exact same queries work immediately via the CLI.

Reproduction Steps

1. Configure Semble as MCP server in the agent config:

   {
     "mcp": {
       "semble": {
         "type": "local",
         "command": ["uvx", "--from", "semble[mcp]", "semble"]
       }
     }
   }

2. Start agent session and invoke semble_search with any query:

• Query: JDBC type conversion ResultSet PreparedStatement DataType
• Mode: hybrid or bm25
• top_k: 1–5
• repo:

3. Result: MCP error -32001: Request timed out (repeated across 5+ attempts)
4. Run the same query via CLI:
   uvx --from "semble[mcp]" semble search "AbstractJdbcRowDataConverter"
   Result: Returns results in ~1 second.
   Expected Behavior
   MCP tool should return search results within a reasonable timeout, matching CLI performance.
   Actual Behavior
   MCP tool times out with MCP error -32001: Request timed out on every attempt. CLI works fine with same repo and query.
   Additional Context

• CLI version works correctly after a manual semble search run
• No ~/.cache/semble directory exists (first run)
• MCP server appears to start successfully (no startup errors)
• Timeout occurs even with minimal query (DataType, bm25 mode, top_k: 1)
• After CLI indexing, MCP tool eventually works on subsequent calls
• Suggests the MCP server may not be handling initial indexing asynchronously or may have a different timeout/initialization path than CLI
  Possible Causes

1. MCP server initialization blocks on indexing without proper async handling
2. MCP transport layer timeout is too short for cold-start indexing
3. Index cache path differs between CLI and MCP server (~/.cache/semble not created)
4. MCP server may not be reusing the same index location as CLI
   Workaround
   Run semble search via CLI once to pre-build the index, then MCP tool works on subsequent calls.

[Pringled]

Hey @cycle2zhou, thanks for reporting and for writing up this detailed issue. The issue here is very likely that the first MCP call has to do multiple things: install semble, download our model, download the relevant tree-sitter language packs, and index your repo. My suspicion is that this leads to timeouts, which would explain why it does work after you've ran the CLI since that won't timeout and after that everything is cached. We're working on a PR to make the model download non blocking now, hopefully that fixes it, will keep you in the loop here.

[Pringled]

@cycle2zhou there's an initial fix released for this which I hope fixes it for you, if possible could you upgrade and try again? Feel free to re-open this if it didn't fix it then I'll dig deeper.