v0.1.0 — Initial Release

sub-mcp — zero-overhead MCP proxy for AI coding agents

Give subagents MCP access without loading schemas into the main session. The main agent's context window stays clean; subagents get full MCP tools on demand.

Features

• Transparent wrapper — installs to ~/.sub-mcp/bin/claude, intercepts claude -p calls and injects --mcp-config
• Seamless tool approval — automatically adds --allowedTools to pre-approve injected MCP tools, so subagents work without interactive permission prompts
• MCP-first system prompt — auto-injects a default system prompt nudging subagents to prefer MCP tools over built-in alternatives (WebFetch, WebSearch). Override per-profile with a custom systemPrompt in profiles.yml.
• Profile-based routing — named MCP configurations with per-profile model, budget, and system prompt overrides
• isolateMcp support — profiles can use --strict-mcp-config to prevent inheriting parent MCPs
• Zero secrets on disk — env vars stored as ${VAR} placeholders, resolved at runtime only
• Direct query mode — sub-mcp query -p <profile> <prompt> for explicit profile routing
• CWD-independent config — sub-mcp refresh works from any directory (falls back to ~/.sub-mcp/profiles.yml)
• Anti-recursion — spawner reads saved real-claude-path to avoid re-entering the wrapper

Quick start

npm install && npm run build
node dist/cli.js install
export PATH="$HOME/.sub-mcp/bin:$PATH"

Full changelog

See README for complete documentation.