Cask — Governed AI agent tools from real API traffic

Turn any web API into a governed, agent-ready MCP server. Lockfile-based approval, fail-closed enforcement, self-repairing proposals, and full audit trail. Every tool your AI agent uses is reviewed, signed, and traceable.

The Problem

AI agents need tools. MCP gives them tools. But who governs what those tools can do?

Without governance, agents silently call admin endpoints, charge billing APIs, leak data to third parties, and escalate their own privileges — with no audit trail and no way to detect drift. OpenAI warns about tool-injection risks. Real data exposure incidents are already happening.

Cask is the missing governance layer: local, deterministic, auditable, fail-closed.

Related MCP server: cordon

See It Work (30 seconds)

pip install caskmcp
cask demo

cask demo compiles 8 tools from bundled API traffic, enforces fail-closed lockfile governance, proves deterministic replay parity, and emits evidence artifacts. Exit code 0 means every governance gate held.

Quick Start (5 minutes)

Prerequisites: Python 3.11+

# 1. Initialize cask in your project
cask init

# 2. Capture traffic and compile a governed toolpack
cask mint https://your-app.com -a api.your-app.com

# 3. Review what changed (risk-classified diff)
cask diff --toolpack .caskmcp/toolpacks/*/toolpack.yaml

# 4. Approve tools for use (interactive TUI or CLI)
cask gate allow --all

# 5. Start the governed MCP server
cask serve --toolpack .caskmcp/toolpacks/*/toolpack.yaml

Your AI agent now has governed, auditable access to your API.

How It Works

  Capture ─── Compile ─── Review ─── Approve ─── Serve ─── Verify
    │            │           │          │           │          │
  HAR/OTEL    tools.json   cask diff  lockfile   MCP stdio  contracts
  OpenAPI     policy.yaml            signatures              drift
  Browser     contracts                                      evidence
  WebMCP      scopes                                         repair

Capture real traffic (HAR, OpenTelemetry, OpenAPI specs, live browser, or WebMCP discovery). Compile into deterministic, versioned tool definitions with risk classification and scopes. Review changes with cask diff — every new tool, schema change, or host addition is risk-classified. Approve via Ed25519-signed lockfile entries — explicit human decisions, not silent defaults. Serve through MCP with fail-closed enforcement — unapproved tools never execute. Verify with contracts, drift detection, and evidence bundles. When things break, cask repair proposes fixes.

What Makes Cask Different

Fail-Closed by Default

No lockfile, no runtime. Period. This isn't a suggestion — it's an architectural invariant. Unapproved tools never execute, and there's no way to bypass it.

Self-Repairing Governance

When agents are denied capabilities, cask repair diagnoses the issue from audit logs, drift reports, and verification failures, then proposes classified fixes:

• SAFE — read-only diagnostics (auto-runnable)

• APPROVAL_REQUIRED — grants new capability (needs human review)

• MANUAL — requires investigation or re-capture

Agents can also propose new capabilities via cask propose — proposals are stored as drafts and only promoted to runtime by explicit human approval.

Interactive TUI

Rich terminal UI for tool review — risk-colored tables, wizard flows, typed confirmation for dangerous operations. Run cask with no arguments for a guided menu.

Agent-Aware Introspection

cask inspect runs a read-only Meta MCP server that exposes governance state as AI-consumable tools. Agents can query what they're allowed to do, check policy, and list pending approvals — making them governance-aware.

Deterministic Replay Parity

Same inputs produce identical artifacts, digests, and tool outputs. This is verified, not aspirational — cask demo proves it in 30 seconds.

Full Audit Trail

Every governance decision (ALLOW, DENY, CONFIRM) is logged with structured traces. Every approval is Ed25519-signed. Every verification run produces an evidence bundle with SHA-256 digests.

Traffic Capture

Start where you already are:

All paths converge to the same governed runtime.

Core Commands

Tip: Run cask with no arguments for an interactive guided menu. Use cask --help-all to see all 25+ commands including compliance, bundle, enforce, confirm, and more.

Runtime Enforcement

The MCP server enforces multiple safety layers on every tool call:

• Lockfile approval — only explicitly approved tools execute

• Policy evaluation — priority-ordered rules (allow, deny, confirm, budget, audit)

• Rate limiting — per-minute/per-hour budgets with sliding-window tracking

• Network safety — SSRF protection, metadata endpoint blocking, redirect validation

• Confirmation flow — HMAC-signed out-of-band challenge tokens for sensitive operations

• Redaction — strips auth headers, tokens, PII from all captured data by default

• Dry-run mode — evaluate policy without executing upstream calls

Installation

Prerequisites: Python 3.11+

# Base install (includes offline demo)
pip install caskmcp

# With MCP server support
pip install "caskmcp[mcp]"

# With live browser capture
pip install "caskmcp[playwright]"
python -m playwright install chromium

# Everything
pip install "caskmcp[all]"

MCP Client Config

Generate a config snippet for your AI client:

# For Claude Desktop
cask config --toolpack .caskmcp/toolpacks/*/toolpack.yaml --format json

# For Codex
cask config --toolpack .caskmcp/toolpacks/*/toolpack.yaml --format codex

Or add this to your Claude Desktop config (~/.claude/claude_desktop_config.json):

{
  "mcpServers": {
    "my-api": {
      "command": "cask",
      "args": ["serve", "--toolpack", "/path/to/toolpack.yaml"]
    }
  }
}

Documentation

• User Guide — full command reference and workflows

• Architecture — system design and component specs

• Glossary — key terms and concepts

• Troubleshooting — common issues and fixes

• Known Limitations — runtime and capture caveats

• Publishing — PyPI release process

Contributing

See CONTRIBUTING.md for development setup, TDD policy, and pull request process.

git clone https://github.com/caskmcp/CaskMCP.git
cd CaskMCP/cask
pip install -e ".[dev,packaging-test]"
pytest tests/ -v

License

MIT