Skip to content

Security

Threat model and security controls for the MCP Ollama Manager VS Code extension.

Trust model

The extension runs in the local VS Code workspace with the same privileges as the user. It spawns the mcp-ollama-python MCP server as a child process and forwards Ollama operations through MCP tools.

Anyone who can change VS Code settings or run extension commands can start/stop the server, pull/delete models, and use AI commands.

Controls (v1.0.9+)

Area Control
mcp-ollama.serverHost Validated hostname/IP; rejects URLs with credentials, paths, or blocked metadata hosts
MCP child process env Minimal passthrough (PATH, locale, etc.) — not full process.env; sets OLLAMA_HOST and OLLAMA_ALLOW_REMOTE_HOST=1 for non-local hosts (aligned with mcp-ollama-python v1.0.8+)
Python path Must be python / python3 / py or an absolute path; shell metacharacters rejected
Model names Validated before pull/delete/show/chat/generate/embed
MCP tools Allowlist only (ollama_chat, ollama_generate, … — not ollama_execute)
MCP prompts Allowlist only (explain_code, write_docstring)
MCP resources Allowlist only (ollama://models, ollama://running, ollama://config)
Logging Sensitive keys redacted in extension logs (src/logger.ts)

Settings

Setting Security note
mcp-ollama.serverHost Use LAN/remote values only on trusted networks
mcp-ollama.pythonPath Point to a trusted Python installation only

The Python MCP server has its own controls (host validation, opt-in code execution). See mcp-ollama-python SECURITY.

Reporting

Report issues via GitHub Security Advisories.