mcp-cocktail

A domain-agnostic, Recursive Self-Improvement (RSI) framework for benchmarking, evaluating, and injecting real-time guardrails across competing MCP servers, CLIs, and AI agent tools.

When software vendors take months or years to release official Model Context Protocol (MCP) servers or command-line tools, open-source communities fill the vacuum with unofficial MCPs, wrappers, and CLIs. This fragmentation leads to competing tools ("arms") with varying degrees of stability, silent failures, and trap behavior.

mcp-cocktail provides the machinery to:

• Inject Real-Time PreToolUse Guardrails: Intercept tool calls right before a known trap is sprung (mcp-cocktail check / install-hook).

• Benchmark Multi-Arm Ecosystems: Run subagents in unison across all available MCPs and CLIs (mcp-cocktail run).

• Log & Mine Friction: Capture real-time friction notes (mcp-cocktail note) and mine subagent transcripts for failure patterns P1-P5 (mcp-cocktail mine).

• Drive Weakness-Maximizing RSI Loops: Synthesize scorecards (mcp-cocktail scorecard --rsi) and auto-derive Weakest Valid Guardrails (traps.json) based on Bennett (2023).

⚡ Quickstart (30 Seconds)

# 1. Install mcp-cocktail
pip install -e .

# 2. Automated Setup & Doctor Health Check for your domain (e.g. unity)
mcp-cocktail setup --preset unity

# 3. Test Guardrail Execution (< 5ms interception)
mcp-cocktail check --selftest

Related MCP server: SentinelMCP

🤝 Zero-Friction Collaborator Setup (Committing Rules to Git)

By committing mcp-cocktail.json, traps.json, and harness configuration files (.claude/settings.json, .omp/settings.json, mcp.json) directly to your project's Git repository, every teammate and AI agent on your team inherits the guardrails automatically:

# Inside your project repo:
git add mcp-cocktail.json traps.json .claude/ .omp/ docs/
git commit -m "chore: add mcp-cocktail guardrails and tool manifest"
git push

When a collaborator clones or pulls the repository:

• Their agent harness (Claude Code, Oh My Pi, Cursor, VS Code) automatically detects the guardrails and configuration files on startup.

• Teammates enjoy real-time trap shielding without needing any manual setup!

📖 Progressive Walkthrough

1. Initialize or Discover Workspace Manifests

Create a workspace config (mcp-cocktail.json) manually or auto-discover candidate MCPs and CLIs from GitHub and registries:

# Auto-discover open-source MCPs & CLIs for a domain (non-destructive merge)
mcp-cocktail discover --domain postgres

# Generate an agentic scout subagent task for deep web discovery
mcp-cocktail discover --domain unity --agentic

mcp-cocktail.json structure:

{
  "name": "unity-ecosystem",
  "description": "Multi-arm evaluation manifest for Unity CLI and MCP servers",
  "arms": [
    {
      "id": "unity-cli",
      "name": "Official Unity CLI",
      "type": "cli",
      "command": "unity",
      "health_check": "unity status --json"
    },
    {
      "id": "official-mcp",
      "name": "Official Unity MCP",
      "type": "mcp",
      "mcp_server": "unity-editor-mcp",
      "tool_prefix": "mcp__unity-editor-mcp__",
      "health_check": "unity status --json"
    },
    {
      "id": "coplay-mcp",
      "name": "CoplayDev Unity MCP",
      "type": "mcp",
      "mcp_server": "UnityMCP",
      "tool_prefix": "mcp__UnityMCP__",
      "setup_script": "tools/three-way-setup.sh"
    }
  ],
  "trial_defaults": {
    "concurrency": "serial",
    "scene_strategy": "auto",
    "timeout_seconds": 300
  }
}

2. Validate Arm Health (mcp-cocktail doctor)

Probe CLI binary PATHs, stdio MCP initialize capabilities, and HTTP endpoints to report an honest status summary:

mcp-cocktail doctor

Reports:

• 🟢 [READY] (CLI active or stdio MCP server initialized with tool count)

• 🟡 [BOUND_ONLY (P4)] (Socket bound but session unauthenticated / unregistered)

• 🟠 [UNCONFIGURED] (Setup script missing or parameter unconfigured)

• 🔴 [OFFLINE] (Process or port unreachable)

3. Native MCP Server Interface (mcp-cocktail serve) & Transparent Proxy (mcp-cocktail proxy)

Mount mcp-cocktail directly into your agent's MCP config (.claude/settings.json or mcp.json) to expose native tools or wrap target MCP servers:

{
  "mcpServers": {
    "mcp-cocktail": {
      "command": "mcp-cocktail",
      "args": ["serve"]
    },
    "unity-editor-mcp": {
      "command": "mcp-cocktail",
      "args": ["proxy", "--", "unity-editor-mcp"]
    }
  }
}

Exposes:

• mcp__mcp-cocktail-server__note_friction

• mcp__mcp-cocktail-server__check_guardrail

• mcp__mcp-cocktail-server__get_scorecard

• mcp__mcp-cocktail-server__run_trial

4. Generate Multi-Arm Trial Briefs & Subagent Payloads

Create standardized briefs and subagent task payloads for subagents to execute the same task independently across defined arms:

# Standard serial trial run (< 1s instant baseline scene reload)
mcp-cocktail run T-001 "Build scene hierarchy for vehicle physics" --exec auto

# Visual comparison mode (leaves temporary scene files for human Unity Editor review)
mcp-cocktail run T-001 "Build scene hierarchy" --compare-visual

Generates:

• docs/trials/T-001/brief-unity-cli.md

• docs/trials/T-001/brief-official-mcp.md

• docs/trials/T-001/trial-meta.json

• docs/trials/T-001/trial-tasks.json

5. Log Friction & Mine Transcripts for P1-P5 Patterns

Subagents or humans can capture friction observations mid-task:

mcp-cocktail note "CLI command ignores positional table filter silently" --cost 15

Mine session transcripts to rank tool usage, identify error clusters, and auto-detect recurring trap patterns (P1 Startup Snapshot, P2 Confident Wrong Answer, P3 Termination $\neq$ Completion, P4 Green Light, P5 Ignored Arguments):

mcp-cocktail mine sweep
mcp-cocktail mine stats <session-uuid>

🔄 The 4 RSI Exhaust Pipelines

When an agent encounters a bug, silent failure, or trap in a tool arm, mcp-cocktail generates 4 distinct, purpose-built exhaust deliverables:

🔬 Theoretical Foundation: Weakness Maximization (Bennett, 2023)

Standard AI theory often relies on Ockham’s Razor or Minimum Description Length (MDL) — assuming that the shortest hypothesis is the most likely to generalize.

As proven by Michael Timothy Bennett (2023) in "The Optimal Choice of Hypothesis Is the Weakest, Not the Shortest" (arXiv:2301.12987v4):

Compression/length is neither necessary nor sufficient for generalization. Instead, to maximize the probability that an inferred hypothesis generalizes, it is necessary and sufficient to select the WEAKEST valid hypothesis — the explanation with maximum generality (least specificity) that remains consistent with observations.

How mcp-cocktail Applies Weakness Maximization:

• An over-fitted guardrail rule (e.g. matching unity command --foo --bar --baz) fails to protect agents calling unity command --other.

• mcp-cocktail uses the Rule of Least Specificity in mcp_cocktail.weakness: it generalizes raw friction observations into the broadest valid regex matchers that maximize coverage over potential tool call spaces while maintaining zero false positives on safe/read-only calls.

mcp-cocktail scorecard --rsi

📁 Architecture & Data Layout

my-project/
├── mcp-cocktail.json              # 1. Manifest: Arms, health checks, CLI commands, capabilities
├── traps.json                     # 2. Rule Store: Active PreToolUse trap rules & matchers
├── .claude/
│   └── settings.json              # 3. Client Config: PreToolUse hook pointing to mcp-cocktail check
└── docs/
    ├── findings-inbox.md          # 4. Raw Friction Inbox: Mid-task append-only notes
    ├── tooling-scorecard.md       # 5. Synthesized Scorecard: Automated ranking table
    ├── upstream/                  # 6. Upstream Vendor Bug Reports: Feedback drafts for official tools
    └── trials/                    # 7. Benchmark Data Store: Trial briefs, meta, tasks, and reports
        └── T-001/
            ├── brief-unity-cli.md
            ├── unity-cli.md
            ├── trial-meta.json
            └── trial-tasks.json

🎮 Reference Datasets

See examples/unity/ for a complete reference dataset containing:

• Multi-arm evaluation manifest (examples/unity/cocktail.json) with 11 curated arms

• Comprehensive Unity trap rule store (examples/unity/traps.json)

• Historic trial reports and scorecard (examples/unity/docs/trials/)

• Historical research log (examples/unity/UNITY-TOOLING-NOTES.md)