Agent Plugin

Portable Agent Plugins package with skills and a gated MCP for ops and coding agents

tokenmaxxing ships a portable Agent Plugins 1.0 package at agent-plugin/. Compatible clients load that directory as a local plugin. Skills under skills/. MCP from mcp.json. Cursor, Codex, ChatGPT, VS Code, Copilot, and friends.

Install

Point the client at agent-plugin/ in a git checkout or an installed npm package (node_modules/tokenmaxxing/agent-plugin).

  • Cursor: load ./agent-plugin as an Agent Plugin, or use the project .cursor/mcp.json stdio entry that runs bun ./agent-plugin/bin/tokenmaxxing-mcp.
  • Pi: project .pi/settings.json already lists agentPluginPaths: ["./agent-plugin"].
  • Other Agent Plugins clients: add the same directory as a local plugin path.

The MCP binary is agent-plugin/bin/tokenmaxxing-mcp (Bun). It refuses ambient CLAUDE_CONFIG_DIR / CLAUDE_SECURESTORAGE_CONFIG_DIR. Set TOKENMAXXING_HOME for hermetic use.

MCP tools

Always available (read / free):

ToolBacking
pool_lstokenmaxxing ls
pool_statustokenmaxxing status (never --force)
doctortokenmaxxing doctor
config_gettokenmaxxing config / config get
helptool catalog + safety rules

Mutating (dual-gated): each call needs confirm: true and the server env TOKENMAXXING_AGENT_MUTATIONS=1. Set that only after the user approves pool mutations.

ToolBacking
pool_switchClaude switch or Codex switch --codex when codex=true
pool_checktokenmaxxing check
config_set / config_unsetconfig writers

Not exposed: status --force, credential dumps, init / add / auth / rm / uninstall, process kills.

Tool results scrub bearer-looking and token-shaped spans. Prefer MCP over raw shell for pool ops.

Skills

SkillJob
pool-statusSafe pool reads
doctor-diagnosticsInstall health
switching-policyGREEDY vs HARD, pace pressure, Layer 2 Claude-only
sdk-pairingAgent SDK helpers
codex-poolRestart-is-switch, no Layer 2
credentials-hygieneIdentity and secret rules
safe-contributionTests, PR window, public-repo hygiene

Skills tell agents when to call which MCP tool and what never to do. Coding agents still edit the repo with normal tools. Policy skills stop banned patterns from coming back.

On this page