Skip to content

ModelBound/modelbound-claude-code-plugin

Repository files navigation

ModelBound — Claude Code plugin

ModelBound Skill Trust

The official ModelBound plugin for Claude Code — keep your team's skills, rules, and system prompts in sync; audit token cost; harden MCP/bash tool use; and run the Skill Development Pipeline without leaving the terminal.

Why ModelBound?

Your skills, rules, and system prompts live in one place — ModelBound — and follow you wherever you work. Write a skill in Cursor, refine it in Claude Code, use it in VS Code with Copilot. No re-writing, no copy-paste between tools. When you switch editors or try a new AI assistant, your context is already there.

This means you get more out of every AI subscription you pay for. Instead of rebuilding your setup from scratch each time you move between tools, ModelBound keeps your library portable. One investment in good context pays off across every platform you touch.

For teams, it's the same story at scale: curate once, distribute everywhere. Everyone stays on the same page regardless of which editor or AI tool they prefer.

Install

# In Claude Code
/plugin install ModelBound/modelbound-claude-code-plugin
mb-login    # one-time device-code auth

Requires Node ≥ 20 and either MODELBOUND_API_KEY or a one-time /mb login.

Slash commands

Skill Development Pipeline (Test & Optimize)

Command What it does
/mb:pipeline <skill-id> Run test → benchmark → optimize pipeline on a skill
/mb:pipeline <skill-id> --dry-run Preview pipeline stages and estimated token cost
/mb:test [skill-id] Run skill tests; omit skill-id to list recent test runs
/mb:versions <skill-id> List saved checkpoints with scores and labels
/mb:restore <skill-id> <version-id> Restore a skill to a specific checkpoint
/mb:diff <skill-id> [from] [to] Diff between two versions (defaults: latest vs current)
/mb:health Local .claude/ token count + remote health scores and budgets

Security

Command What it does
/mb-optimize <file|skill> Run token optimization. Add --apply to save a new version.
/mb-pipeline <skill> Full Skill Development Pipeline (lint → trust → test → benchmark → optimize).
/mb-test <skill> Run the test suite (optional --model).
/mb-benchmark <skill> <verA> <verB> Head-to-head benchmark.
/mb-versions <skill> List versions, newest first.
/mb-restore <skill> <versionId> Restore (non-destructive — creates a new version).
/mb-diff <skill> <from> [to] Unified diff between versions.
/mb-health Check API connectivity + auth.
/mb-login / /mb-logout / /mb-whoami Auth shortcuts.

Pre-edit backup hook

  • SessionStart — runs /mb:sync-rules if you opted in; warns on drift
  • PostToolUse(Edit) on .claude/** — auto-pushes edits to ModelBound
  • PreToolUse(Edit|Write|MultiEdit) — snapshots files to .mb-backup/ before editing (best-effort)
  • PreToolUse(Bash) — blocks a configurable denylist (rm -rf, curl | sh)
  • PreToolUse(WebFetch) — blocks private IP ranges and non-allow-listed domains

Disable any hook in ~/.modelbound/config.json:

{ "hooks": { "autoSync": false, "bashGuard": true, "webFetchGuard": true } }

Subagents

  • mb-reviewer — reviews diffs using your team's ReviewPanel rubric
  • mb-context-doctor — diagnoses CLAUDE.md bloat, suggests Smart Split

Config

Config lives at ~/.modelbound/config.json (created on sign-in):

{
  "apiKey": "mb_live_...",
  "activeTeamId": "uuid",
  "mcpUrl": "https://mcp.modelbound.co/mcp",
  "authUrl": "https://modelbound.co/api/extension-device-auth"
}

Updating

claude plugin update modelbound

Disable per-session with MODELBOUND_DISABLE_BACKUP=1.

Related projects

Project Description
ModelBound CLI Terminal + CI for token optimization, skill pipeline, and version management
ModelBound MCP Server · npm Local-first MCP server for skill lint, convert, and cloud sync
Cursor Extension · Marketplace VS Code/Cursor extension for rules sync and MCP bridge
Cursor Plugin Cursor slash commands for pipeline, trust & safety, and versions
Dev Packs Open-source curated AI context packs for engineering teams

Install hub: modelbound.co/connect

License

MIT