v0.3.0 — Drop deprecated kimi-k2-thinking-turbo

Breaking

• Removed kimi-k2-thinking-turbo (deprecated upstream by Moonshot). The aliases kimi and kimi-k2 now resolve to kimi-k2.6, so callers using model: kimi continue to work but now hit the multimodal K2.6 endpoint (vision-capable, temperature=1.0 enforced via FixedTemperatureConstraint). Callers that depended on the wider temperature range (0.0–1.0) on the old turbo model will have temperature clamped to 1.0.

---

Install: uv tool install vox-mcp (or pipx install vox-mcp)
Full changelog: https://github.com/linxule/vox-mcp/blob/main/CHANGELOG.md

v0.2.0 — Model swap: Kimi K2.6 + DeepSeek V4 Pro

Breaking

• Removed kimi-k2.5 (and its aliases k2.5, kimi-k25). Replaced by kimi-k2.6 (256K context, 64K output, multimodal with vision, always-on thinking). Aliases: k2.6, kimi-k26. Thinking mode requires temperature=1.0 (now enforced via FixedTemperatureConstraint).
• Removed deepseek-reasoner (and its aliases deepseek-r1, ds-reasoner, r1). Replaced by deepseek-v4-pro (1M context, 384K output, always-on thinking). Aliases: deepseek, deepseek-v4, v4, v4-pro. Provider now injects extra_body={"thinking": {"type": "enabled"}} since V4 Pro is a single endpoint with a thinking toggle, unlike R1.

Notes

• OpenRouter routing is unaffected; R1 references in _TEMP_UNSUPPORTED_PATTERNS retained for OpenRouter's deepseek/deepseek-r1-* models.

---

Backfilled retroactively. Full changelog: https://github.com/linxule/vox-mcp/blob/main/CHANGELOG.md

v0.1.0 — Initial public release

First public release of Vox MCP.

Features

• 8 providers: Google Gemini, OpenAI, Anthropic, xAI (Grok), DeepSeek, Moonshot (Kimi), OpenRouter, and custom OpenAI-compatible endpoints (Ollama, vLLM, LM Studio)
• 3 tools: chat (multi-model gateway with file/image context), listmodels (model discovery), dump_threads (thread export as JSON or Markdown)
• Conversation memory: in-memory threads with continuation_id for multi-turn exchanges across any provider
• Thread persistence: append-only JSONL with cold-reload from disk
• Thinking mode support: per-model constraint architecture (token budgets, effort levels, always-on reasoning)
• Temperature constraints: fixed, range, and discrete handling per model
• Model restrictions: per-provider allowlists via environment variables
• Auto mode: agent-driven model selection when DEFAULT_MODEL=auto

Design

Pure API passthrough — no system prompt injection, no response modification. Conversation memory is the sole value-add.

Install

git clone https://github.com/linxule/vox-mcp.git
cd vox-mcp
cp .env.example .env  # add at least one API key
uv sync
uv run python server.py

Also available on PyPI and the MCP Registry.

Attribution

Derived from pal-mcp-server by Beehive Innovations. Apache 2.0.