activeing123/mcptoon

pip install mcptoon

Usage

• •Add servers one command each: mcptoon add fetch --stdio npx -y @modelcontextprotocol/server-fetch.
• •List tools without polluting context: mcptoon manifest --compact (or --slim / --toon).
• •Call a tool: mcptoon call fetch fetch '{"url":"https://example.com"}' --toon.
• •Agents share one config (~/.mcptoon/config.json) across all agents; use mcptoon quickstart for one-command onboarding.

Key features

• ✓Tool schemas never enter the agent context — compact mode is 98.5% smaller than JSON (tiktoken-verified).
• ✓One config file shared by every shell-capable agent; delete mcptoon and your servers keep running.
• ✓Built-in dangerous-op, prompt-injection, and credential-leak guards.
• ✓Zero third-party imports (~200KB), no telemetry, no credential storage.
• ✓Python API with a round-trip-safe TOON codec; auto-discovery via quickstart/init.

Best for

Users running many MCP servers across multiple shell-capable agents who want to stop context-window bloat.

Caveats

• ⚠NOT an MCP server — it does not plug into mcpServers JSON and works only with shell-capable agents (Claude Code, Codex, Cursor, OpenCode, etc.).
• ⚠Does not replace native MCP config in Cursor or Claude Desktop, and exposes no MCP JSON-RPC endpoint.
• ⚠--fallback-json only catches encoding-level errors, not whether the LLM parsed the output correctly.
• ⚠Destructive ops (delete/drop/purge) are blocked by default unless --destructive is passed.