0.3.43
Fixes a broken MCP install
mcp 2.0.0 removed mcp.server.fastmcp, which the MCP server imports at module scope. The extras allowed it through an unbounded mcp>=1.10.0, so any fresh install since mcp 2.0.0 shipped produced a server that would not start:
$ pip install 'llm-sandbox[mcp-docker]'
ModuleNotFoundError: No module named 'mcp.server.fastmcp'
This affected everyone installing 0.3.42 fresh, including via the MCP registry entry. Upgrade if you use the MCP server.
What changed
mcp>=1.10.0 → mcp>=1.28.1,<2 across the mcp-docker, mcp-podman, and mcp-k8s extras.
The ceiling stops 2.0.0. The floor is not arbitrary — it also clears two high-severity advisories in mcp:
| Advisory | Fixed in |
|---|---|
| WebSocket transport lacks Host/Origin validation | 1.28.1 |
| HTTP transports serve sessions without verifying the authenticated principal | 1.27.2 |
Neither was exploitable here, since this server runs over stdio, but there is no reason to keep permitting the vulnerable range.
Resolves to mcp 1.29.0. Support for mcp 2.x requires migrating off mcp.server.fastmcp and will follow separately.
Also in this release
Repository and metadata only, no runtime impact: CITATION.cff now updates automatically on release, and the JOSS paper builds in CI.
Full changelog: 0.3.42...0.3.43