ModelBound MCP server
Commit 18. Okt. 2026Syncs Cursor rules, Claude Code skills, and system prompts across IDEs and provides an MCP bridge for AI agent context management.
Auch bekannt als ModelBound/modelbound-cursor-extension
- Stars
- 0
- Tools
- —
- Downloads, 30 Tage
- —
- Neueste Version
- —
- Letzter Commit
- 18. Okt. 2026
- Lizenz
- MIT
Was ist ModelBound MCP?
Syncs Cursor rules, Claude Code skills, and system prompts across IDEs and provides an MCP bridge for AI agent context management.
Zusammenfassung generiert aus dem README des Projekts
Über ModelBound MCP server
Synchronizing AI instructions across different development environments is what ModelBound enables for engineers using Cursor, Claude Code, or Copilot. This extension acts as a local bridge that watches specific directories like .cursor/rules or .claude to ensure your system prompts and skills remain consistent across multiple tools. It allows you to write a skill once and have it automatically available regardless of which AI agent or editor you are currently using.
Users can manage their prompt libraries through a local-first approach that keeps source code private. The system provides a skill development pipeline to test, optimize, and deploy prompt changes via the command palette. While basic token footprint estimation runs locally without any account, full cloud synchronization and version control require signing in via OAuth or an API key. The ModelBound MCP server complements this by exposing over 40 tools to compatible agents for fetching skills and running evaluations.
Was du damit machen kannst
- 1Syncing rules across IDEsMaintain consistent AI behavior when switching between Cursor and Claude Code on the same project.
- 2Managing team prompt librariesProvide a single source of truth for system prompts and rules across an entire organization.
- 3Validating prompt changesRun a CI/CD-style pipeline to test and optimize AI skills before they go into production.
- 4Monitoring instruction sizeCheck the estimated token footprint of your instruction files directly in the status bar.
Geschrieben aus dem README des Projekts. Jeder Use Case wurde gegen ein Zitat daraus geprüft.
Automatisierte Checks
Geluften am 15. Nov. 2026Vollständiger Bericht →Noch nicht verifiziert. Ein Server benötigt einen Score von 80, keine kritischen Befunde und genug Daten für unsere Prüfung. Wie die Checks funktionieren · Was du für mehr Punkte verbessern kannst
- Source Repository verknüpft und erreichbarmodelbound/modelbound-cursor-extension is public and reachablegithub.com/modelbound/modelbound-cursor-extension ↗Bestanden
- Paket nennt dieses Repository als QuelleNicht geprüft
- Build-Provenienz (Sigstore-Attestierung)Nicht geprüft
- Open Source Lizenz (OSI)MITspdx.org/licenses/MIT.html ↗Bestanden
- Commit oder Release im letzten JahrLast commit or release 2026-10-18github.com/modelbound/modelbound-cursor-extension ↗Bestanden
- OpenSSF ScorecardNicht geprüft
- Package-Name ähnelt keinem populären PackageNicht geprüft
- Keine kritischen Advisories im Paket oder seinen AbhängigkeitenNicht geprüft
- Keine versteckten Anweisungen an das Modell0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersBestanden
- Destruktive Tools sind markiertNicht geprüft
- Remote-Endpoint beantwortet den MCP-HandshakeNicht geprüft
README
Auf GitHub ansehenModelBound — Cursor Rules Sync, AI Skills Manager & MCP Prompt Server
The system prompt version control layer for your IDE. Keep your Cursor rules, Claude Code skills, .cursorrules, and AI knowledge bases in sync between your workspace and ModelBound.co — automatically. The ModelBound MCP server gives Cursor, Claude Code, Copilot, and any MCP client live access to your shared prompts, eval results, and token-health insights.
Open source. Local-first. Zero source code leaves your machine.
Instant instruction profile — no account needed
Open a SKILL.md, CLAUDE.md, AGENTS.md, Cursor rule, or Copilot instruction file. The status bar shows its estimated token footprint and updates as you type, including unsaved changes. Click it for safety findings and jump to the relevant line. This runs locally with no AI calls. It estimates instruction size, not actual runtime tokens or cost. Sign in only when you want cloud sync and real-run insights.
Why ModelBound?
If you've ever dealt with:
.cursorrulesfiles drifting out of sync across repos- Manually copying system prompts between Cursor, Claude, and Copilot
- No version history for your AI rules and context
- Wanting an MCP prompt server that actually works with your IDE
- No CI/CD pipeline for testing and validating prompt changes
- Rebuilding your entire prompt library when switching AI tools
This extension solves all of it.
One Library, Every AI Tool
Your prompts, rules, and skills shouldn't be locked into a single IDE or agent. ModelBound acts as the shared layer underneath — write a skill once, and it's available everywhere you work:
- Switch from Cursor to Claude Code mid-project? Your context follows you.
- Try a new AI tool next month? Your entire prompt library is already there.
- Run Copilot at work and Cursor at home? Same skills, same quality, no re-work.
Stop rebuilding context from scratch every time you change tools. ModelBound makes your investment in prompt engineering portable — so you get full value from every AI subscription you're paying for, regardless of which editor or agent you're using today.
Getting Started
1. Install
# Cursor
cursor --install-extension ModelBound.modelbound-cursor-extension
# VS Code
code --install-extension ModelBound.modelbound-cursor-extension
Or search "ModelBound" in the Cursor/VS Code Extensions panel, or download the .vsix from GitHub Releases.
2. Sign In
On first activation, you'll be prompted to authenticate. Two options:
- Sign In with Browser (recommended): Opens your browser for a one-click OAuth sign-in. No tokens to copy — the extension handles everything automatically via the Device Authorization Grant flow.
- Paste API Key: For users who prefer manual setup or CI environments. Grab your key from modelbound.co/settings and paste it when prompted.
You can also sign in later via Cmd+Shift+P → "ModelBound: Sign In with Browser".
3. Start Editing
That's it. Save any file in a watched context directory and it syncs to ModelBound cloud automatically. Pull skills from your team's shared library with Cmd+Shift+P → "ModelBound: Pull Skill/Context to Local Workspace".
Features
- One-Click OAuth Sign-In: Browser-based authentication — no manual token management. The extension opens your browser, you approve, and you're connected.
- Cursor Rules Sync: Edit
.cursor/rules/files natively. On save, changes push to ModelBound cloud and stay versioned. .cursorrulesManager: Pull shared rules from your team's ModelBound workspace directly into your local Cursor config.- Claude Code Skills Sync: Same bi-directional sync for
.claude/— your Claude skills stay current across machines. - System Prompt Version Control: Every edit is tracked. Roll back, branch, or share prompts across your team from the ModelBound dashboard.
- Skill Development Pipeline: Run test, optimize, and production stages against your skills directly from the IDE. Visual status panel shows pass/fail results in real time.
- MCP Prompt Server: The ModelBound MCP server exposes 40+ tools to any MCP-compatible agent. This extension handles the local filesystem bridge — writing files and syncing edits back to the cloud.
- IDE-Aware File Placement: Automatically writes pulled skills to the correct native location for your IDE —
.kiro/skills/,.cursor/rules/,.claude/, and.modelbound/. - Git-Aware Sync: Detects your repo URL and branch, associating synced skills with the correct project context server-side.
- Local Isolation: Only context directories are watched. No application source code, credentials, or telemetry ever leave your machine.
How It Works
+--------------------+ +---------------------------+ +---------------------+
| ModelBound Cloud | <===========> | This Extension (Local) | <===========> | IDE / Agent |
| (Context Hub) | Secure API | File Watcher + Sync | Filesystem | (Cursor, Kiro, etc) |
+--------------------+ +---------------------------+ +---------------------+
|
v
Watches & syncs these directories:
• .modelbound/**/*.{md,json}
• .kiro/skills/**/*.md
• .cursor/rules/**/*.md
• .claude/**/*.md
• .agents/skills/**/SKILL.md
The ModelBound MCP server runs remotely and handles agent-facing tools (fetching skills, searching context, running evals). Since it can't see your local filesystem, this extension acts as the local bridge — writing files to the right places and syncing edits back.
Supported IDEs & Tools
| IDE / Tool | Context Directory | What Syncs |
|---|---|---|
| Cursor | .cursor/rules/ |
Cursor rules, .cursorrules |
| Kiro | .kiro/skills/ |
Skills and steering files |
| Claude Code | .claude/ |
Claude skills and context |
| Agents | .agents/skills/ |
SKILL.md files |
| Copilot | .github/ |
Copilot instructions (coming soon) |
| Any MCP client | .modelbound/ |
All skills (canonical) |
The extension never creates IDE parent directories from scratch — it only writes into them if they already exist in your workspace.
Commands
| Command | Description |
|---|---|
ModelBound: Sign In with Browser |
OAuth sign-in — opens browser, no token copy needed |
ModelBound: Sign Out |
Clear stored credentials |
ModelBound: Pull Skill/Context to Local Workspace |
Fetch a skill by ID and write to all detected IDE locations |
ModelBound: Set API Key (Manual) |
Manually set or update your API key |
ModelBound: Run Skill Development Pipeline |
Run test/optimize/production pipeline against a skill |
Skill Development Pipeline
Run CI/CD-style validation on your AI skills without leaving the IDE:
Cmd+Shift+P→ "ModelBound: Run Skill Development Pipeline"- Select a stage: Full pipeline, Test & Optimize only, or Production only
- Choose production targets: Save, Publish to Marketplace, Export to Claude
- A live status panel opens showing pass/fail results for each stage
The pipeline runs server-side via the ModelBound MCP server and reports results back in real time.
Use Cases
- Team prompt libraries: Share system prompts and Cursor rules across your org. Everyone pulls from the same source of truth.
- Multi-IDE workflows: Use Cursor and Claude Code on the same project? Edits in either sync back to ModelBound and propagate to both.
- Tool-agnostic prompt investment: Build your skills once in ModelBound and use them across every AI tool you subscribe to — no vendor lock-in, no re-work when you switch.
- Prompt version control: Track changes to your AI rules over time. See who changed what and roll back if needed.
- Skill CI/CD: Validate prompt changes before publishing — run test gates, optimization passes, and production deploys from the command palette.
- MCP-native agents: Let your AI agent fetch and scaffold context directly via the ModelBound MCP server — no manual file management.
Extension Settings
| Setting | Type | Default | Description |
|---|---|---|---|
modelbound.apiKey |
string | "" |
Your ModelBound.co API Key (set automatically via sign-in) |
modelbound.autoSync |
boolean | true |
Auto-sync context file changes back to cloud |
modelbound.mcpUrl |
string | https://mcp.modelbound.co |
MCP Streamable HTTP endpoint |
modelbound.authUrl |
string | https://modelbound.co/api/extension-device-auth |
Device auth endpoint (advanced) |
Privacy & Security
This extension is open-sourced so enterprise security teams can audit exactly what data moves off their machines:
- Directory Isolation: The file watcher only targets known context directories listed above. It cannot index outside them.
- Exclusion Manifest: A
.modelboundignorefile prevents accidental tracking of source code or secrets. - No telemetry. No analytics. Just file sync.
- OAuth tokens stay local: API keys are stored in your VS Code/Cursor settings — never transmitted anywhere except the ModelBound API.
Related projects
| Project | Description |
|---|---|
| ModelBound CLI · npm | 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 Plugin | Cursor slash commands for pipeline, safety checks, and versions |
| Claude Code Plugin | Claude Code plugin for pipeline, hooks, and skill sync |
| Dev Packs | Open-source curated AI context packs for engineering teams |
Also on Open VSX. Install hub: modelbound.co/connect
Contributing
- Clone and install:
git clone https://github.com/ModelBound/modelbound-cursor-extension.git
cd modelbound-cursor-extension
npm install
- Watch for changes:
npm run watch
- Press
F5to launch the Extension Development Host for testing.
License
Distributed under the terms of the MIT License.
Feedback loop
| Command | What it does |
|---|---|
| ModelBound: Report Skill Outcome… | Record worked / partly worked / went wrong for the active skill, with a failure category, optional note, and selected output excerpt. |
| ModelBound: Show Skill Reliability | Show per-skill reliability for the last 30 days. |
Outcome reports measure skill quality after real use. Repeated failures are grouped and diagnosed in ModelBound, where a minimal fix and regression test can be reviewed.
The agent harness
The pipeline panel now shows a fourth check alongside readiness checks, latency and tests: Safety — a deterministic read on whether the skill is cleared to run unattended, across context, permissions, guardrails, verification. It arrives with the existing pipeline result, so there is nothing to configure.
A skill that fails the gate is labelled supervised only, with the blocking
reasons and a suggested fix for each. Presets for the three common postures live
in presets/harness.json.
Tracing
Runs reported through the MCP server (report_run), the CLI (mb trace)
or any OpenTelemetry exporter appear under Live runs on the skill page:
a per-version scorecard and a step-by-step timeline. Summaries only; see
docs/TRACING.md.
Tests
npm run test:validate # package.json ↔ registerCommand parity
npm run test:unit # MCP parse, outcomes vocabulary, harness presets
npm test # VS Code extension host (CI uses xvfb-run)
CLI vs extension vs plugin coverage: docs/PARITY.md.
ModelBound MCP Fragen
Does my source code leave my machine?
No, the extension uses local isolation where only specific context directories are watched and no application source code or credentials leave your machine.
How do I sync my prompts across different AI tools?
ModelBound acts as a shared layer that allows you to write a skill once and have it available in Cursor, Claude Code, Copilot, and other MCP-compatible agents.