DEV Community

MCP Token Saver
MCP Token Saver

Posted on

I Shipped a Plugin That Teaches Your AI Agent to Manage Its Own MCP Config

#mcptoon #ai #mcp

Your agent uses tools every day, but it has no idea how the plumbing works. Ask it why your MCP setup burns 70K tokens of schemas per turn and it will guess. Ask it to wire a new server into five agent configs and it will edit the wrong file, or worse, the right file in the wrong shape.

That is a skills gap, not an intelligence gap. So with mcptoon 0.7.2 we shipped the fix as a plugin: three SKILL.md files that ride along as MCP prompts, teaching any agent how to drive mcptoon — connect, author, triage. mcptoon eats its own dog food here: the skills are installed and served through mcptoon's own plugin system.

The install is now one line

The whole onboarding used to assume you were comfortable with pip and a README walk. Version 0.7.2 collapses it:

macOS / Linux / WSL:

curl -fsSL https://raw.githubusercontent.com/activeing123/mcptoon/main/install.sh | bash
Enter fullscreen mode Exit fullscreen mode

Windows (PowerShell):

irm https://raw.githubusercontent.com/activeing123/mcptoon/main/install.ps1 | iex
Enter fullscreen mode Exit fullscreen mode

The script picks the best available path (pipx, then pip --user, then a PEP 668-safe venv on macOS where the system Python refuses global installs), fixes PATH if needed, and hands off straight to mcptoon quickstart. You end the one-liner staring at a list of your own MCP tools.

Quickstart now ends with the payoff

mcptoon quickstart adopts your existing setup — it scans the agents you already configured and lists every tool it finds, so nothing gets re-entered. In 0.7.2 it closes with the number that matters:

🎉 255 tools ready across 8 servers!

Now you can:
  mcptoon sync      # write configs into every agent
  mcptoon serve     # expose all servers as one stdio gateway
Enter fullscreen mode Exit fullscreen mode

sync writes native-format configs into each detected agent — merging, never overwriting, so your hand-tuned entries survive. serve turns the whole pile into one stdio gateway, which is where the token compression pays off: tool listings drop by up to 99% once agents read the compressed manifest instead of the raw schemas.

The skills plugin: three files, one command 

mcptoon plugin install ~/.mcptoon/plugins/mcptoon-skills
Enter fullscreen mode Exit fullscreen mode

What the agent gets:

  • connect — how to inspect an existing setup, run quickstart/sync, and verify configs landed correctly
  • authoring — how to write the compressed manifest format and when compression is worth it
  • triage — how to debug a failing MCP server: health checks, config drift, restart semantics

They are exposed through mcptoon serve as MCP prompts, so any connected agent can pull them in on demand. No new tool calls, no extra config files — the knowledge arrives through the same channel the tools do.

Try it 

# 1 · one-line install (about 250KB, pure standard library, zero dependencies)
curl -fsSL https://raw.githubusercontent.com/activeing123/mcptoon/main/install.sh | bash

# 2 · see the savings on your own setup before committing to anything
mcptoon demo

# 3 · give your agent the manual
mcptoon plugin install ~/.mcptoon/plugins/mcptoon-skills
Enter fullscreen mode Exit fullscreen mode

The suite behind it runs 610 tests, green in CI. It is a local CLI, not a service — nothing leaves your machine.

Your agent was never missing intelligence. It was missing the manual. Now the manual ships with the toolbox.

Top comments (0)