fix: MCP startup timeout, Pydantic crash, proxy support, and encoding

[bennylii]

Summary

Fixes the remaining issues that prevent the MCP server from working on Windows and in proxied environments. Builds on PR #2 (@OctavioCriollo's structured_output=False approach for the vision tools) and adds several additional fixes.

Issues fixed

1. Pydantic crash on vision tools (same as PR #2)

Added structured_output=False to analyze_image and analyze_image_json so the MCP SDK skips schema generation for the Image return type.

2. MCP startup timeout (~15s → ~2s)

httpx.Client() instantiation on Windows takes ~2 seconds per call due to SSL/certificate store overhead. All three modules (groq_tts.py, groq_stt.py, groq_vision.py) created their clients at module level, adding ~6s+ to import time. Converted to lazy _get_client() pattern — clients are created on first API call instead of at import.

3. Dead code in groq_stt.py

groq_stt.py created a module-level groq_client that was never used — the functions called httpx.post() directly. Removed the dead client. Also migrated httpx.post() calls to use the lazy _get_client() to get proxy support.

4. Proxy support

None of the modules respected HTTP_PROXY / HTTPS_PROXY environment variables. The lazy _get_client() functions now read these env vars and configure the httpx client accordingly.

5. UTF-8 encoding on output files

groq_vision.py used open(file, "w") without encoding — on non-UTF-8 systems (Chinese Windows, etc.) this causes UnicodeEncodeError when writing analysis results. Changed to open(file, "w", encoding="utf-8").

6. Unsafe stdout write

Removed print("Starting Groq TTS server") from main() — printing to stdout is unsafe for MCP stdio transport.

Files changed

File	Changes
server.py	+2 structured_output=False, -1 print
src/groq_tts.py	lazy client + proxy, → _get_client()
src/groq_stt.py	lazy client + proxy, → _get_client(), remove dead groq_client
src/groq_vision.py	lazy client + proxy, → _get_client(), encoding="utf-8"

Tested

• Windows 10, Python 3.12, mcp 1.27.0, pydantic 2.12.5, httpx 0.28.1
• Import time: ~15s → ~2s
• Server starts under Claude Desktop / Claude Code stdio transport
• Vision analysis with analyze_image confirmed working through proxy (HTTP_PROXY)

CC @OctavioCriollo — this extends your PR #2 with additional fixes I found while debugging Windows startup.

[mashkovtsevlx]

Confirming the Pydantic crash reproduces on macOS (Python 3.11, mcp[cli] latest), so it isn't Windows-specific — the server fails at import for everyone on a recent MCP SDK:

pydantic.errors.PydanticSchemaGenerationError: Unable to generate pydantic-core schema for <class 'mcp.server.fastmcp.utilities.types.Image'>

The structured_output=False change in this PR fixes it. As a stopgap for anyone blocked right now, pinning mcp[cli]==1.6.0 also avoids the crash. Would be great to get this merged.

cc @janzheng (maintainer) — thanks @bennylii for the fix.