Rostr

Intelligent MCP proxy for Cursor. One server instead of eight — 90% less context window bloat.

Rostr sits between Cursor's AI and your MCP servers. It auto-discovers your connected servers, learns patterns from every workflow run, and provides optimized sequencing advice — all through 4 lightweight tools instead of flooding the context window with hundreds of tool definitions.

Install

npm install -g rostr-mcp
rostr init

Cursor Setup

Add Rostr to your .cursor/mcp.json:

{
  "mcpServers": {
    "rostr": {
      "command": "npx",
      "args": ["-y", "rostr-mcp"]
    }
  }
}

Restart Cursor. Rostr will auto-discover your other MCP servers (Neon, Stripe, Vercel, etc.) and write a .cursor/rules/rostr.mdc playbook that Cursor reads on every interaction.

How It Works

┌─────────────┐    ┌───────────┐    ┌─────────────────────┐
│  Cursor AI  │───▶│   Rostr   │───▶│  .cursor/rules/     │
│  (4 tools)  │    │  (local)  │    │  rostr.mdc          │
└─────────────┘    └─────┬─────┘    └─────────────────────┘
                         │
                   ┌─────▼─────┐
                   │  ~/.rostr │
                   │  SQLite   │
                   └───────────┘

1. Discover — Rostr reads .cursor/mcp.json and plugin directories to find your servers
2. Advise — suggest_plan returns the optimal step sequence with warnings from past failures
3. Learn — log_run records outcomes, extracts patterns, updates the .mdc playbook
4. Recall — recall_playbook returns failure patterns and success rates for any stack combination

MCP Tools

| Tool | When Cursor uses it | |------|-------------------| | list_roster | Start of any infra task — see connected servers and saved workflows | | suggest_plan | Before multi-step work — get the optimal sequence with failure warnings | | log_run | After completing a task — log the outcome so Rostr learns. Accepts optional task_type (database, deployment, payments, auth, ci-cd) | | recall_playbook | Before unfamiliar stack combos — get known patterns and success rates |

CLI Commands

rostr init                          # Discover servers, create DB, write .mdc
rostr status                        # Show servers, stats, telemetry, pattern count
rostr workflow list                  # List saved workflows
rostr workflow add --name "deploy"   # Add a workflow (--file or --steps)
rostr workflow remove "deploy"       # Remove a workflow
rostr logs                           # Recent runs (--stack, --limit)
rostr patterns                       # Learned patterns (--stack, --limit)
rostr export                         # Export playbook as JSON (--stack to filter)
rostr telemetry on|off|status        # Manage anonymous telemetry opt-in
rostr reset --yes                    # Clear all data

Pattern Learning

Rostr extracts patterns automatically after each logged run:

• Failure rates — "run_migration fails 43% of the time"
• Ordering — "create_branch must complete before run_sql"
• Timing — "fast failures suggest adding delays between steps"
• Success rates — "overall success rate: 60%"

Patterns are stored locally in ~/.rostr/rostr.db and surfaced in the .mdc playbook and through suggest_plan / recall_playbook.

Token Efficiency

Without Rostr, 8 MCP servers expose ~200 tool definitions to the AI, consuming thousands of context tokens on every interaction. Rostr replaces all of that with 4 tools and a lightweight .mdc rule file, keeping pattern context under 200 tokens.

License

MIT