MCPg v0.6.9 — release notes

Released: 2026-07-07 Tool surface: 252 tools across 19 capability buckets (read-only mode exposes ~185) Tests: unit + integration suite green (PG 14 / 15 / 16 / 17 / 18 / 19 / WarehousePG) Runtime: Python 3.14

A patch bump (0.6.8 → 0.6.9) headlined by a big expansion of the natural-language-to-SQL provider fleet, plus two correctness fixes. Backward-compatible — no tool signatures changed.

Headline: 19 built-in NL→SQL providers, plug-and-play

translate_nl_to_sql now ships 19 built-in providers, up from seven. The expanded fleet adds xAI (Grok), GitHub Models, Hugging Face, Groq, Mistral, Together, Fireworks, DeepInfra, Cerebras, Nebius, SambaNova, and Moonshot (Kimi) alongside the original Anthropic, OpenAI, Gemini, DeepSeek, Qwen, OpenRouter, and Perplexity.

Each is genuinely plug-and-play: set the vendor’s conventional API-key env var and MCPg auto-discovers it — including the ones that deviate from the <VENDOR>_API_KEY convention (Hugging Face → HF_TOKEN, GitHub Models → GITHUB_TOKEN, DeepInfra → DEEPINFRA_TOKEN). Every base URL and key env var was verified against the vendor’s own docs.

Under the hood, all provider metadata now lives in a single declarative registry (nl2sql.ProviderSpec / _PROVIDERS) that every lookup table derives from. Adding a provider — or refreshing a default model when a vendor retires one — is a one-line data edit, no code change. Local stacks (Ollama, vLLM, LM Studio) stay first-class on the keyless MCPG_NL2SQL_CUSTOM_PROVIDERS path.

Also fixed

• Windows: the HTTP transport now connects to Postgres. Running MCPg with MCPG_TRANSPORT=streamable-http (or sse) on Windows previously failed every database connection with a 30-second pool timeout — uvicorn reinstalled the ProactorEventLoop, which async psycopg rejects. run_http now pins the WindowsSelectorEventLoopPolicy and serves uvicorn on it. (stdio was unaffected.)
• serverInfo now reports MCPg’s own version. The MCP initialize handshake had been advertising the MCP SDK’s version instead of mcpg’s; the advertised version is now pinned to mcpg.__version__.

Under the hood

• Version single-sourced from src/mcpg/__init__.py. pyproject.toml reads it dynamically via hatchling, so pip, mcpg --version, and the serverInfo handshake can never disagree — a release bumps exactly one line.

Upgrade

pip install --upgrade mcpg
docker pull ghcr.io/devopam/mcpg:0.6.9   # or :latest

Or grab mcpg-0.6.9.mcpb from this release and double-click it into Claude Desktop. No configuration changes required.

Full changelog

See ../CHANGELOG.md [0.6.9] for the complete itemised list.