io.github.VouchlyAI/pincer

Official

by VouchlyAI · TypeScript

Secure grip for your agent's secrets - security-hardened MCP gateway with proxy token architecture

Security-hardened MCP gateway with proxy token architecture for secret management.

Pincer MCP 🦀

Pincer-MCP is a security-hardened Model Context Protocol (MCP) gateway that eliminates the "Lethal Trifecta" vulnerability in agentic AI systems. By acting as a stateless intermediary, Pincer ensures agents never see your real API keys.

🔒 The Problem

Current AI agents store long-lived API keys in plain-text .env files or local databases. If compromised via prompt injection or host intrusion, attackers gain direct access to your:

Database passwords
Third-party API keys

✨ The Solution: Proxy Token Architecture

Pincer implements a "blindfold" security model:

Agent knows: Only a unique proxy token (pxr_abc123...)
Pincer knows: Mapping of proxy tokens → real API keys (encrypted in OS keychain)
Agent never sees: The actual credentials

mermaid

sequenceDiagram
    participant Agent
    participant Pincer
    participant Vault (OS Keychain)
    participant External API

    Agent->>Pincer: tools/call + proxy_token: pxr_abc123
    Pincer->>Vault: Decrypt real API key
    Vault-->>Pincer: gemini_api_key: AIzaSy...
    Pincer->>External API: API call with real key
    External API-->>Pincer: Response
    Pincer->>Pincer: Scrub key from memory
    Pincer-->>Agent: Response (no credentials)

📦 Available Tools

gemini_generate: Secure Google Gemini API calls.
openai_chat: Chat completions with OpenAI GPT models (gpt-4o, gpt-4-turbo, gpt-3.5-turbo, etc.).
openai_list_models: List all available OpenAI models.
openai_compatible_chat: Chat completions with any OpenAI-compatible API (Azure OpenAI, Ollama, vLLM, etc.).
openai_compatible_list_models: List models from custom OpenAI-compatible endpoints.
claude_chat: Chat completions with Anthropic Claude models (Claude 3.5 Sonnet, Opus, Haiku).
openrouter_chat: Unified API access to 100+ models from multiple providers (OpenAI, Anthropic, Google, Meta, etc.).
openrouter_list_models: List all available models across OpenRouter providers.
openwebui_chat: OpenAI-compatible interface for self-hosted LLMs.
openwebui_list_models: Discover available models on an OpenWebUI instance.
gpg_sign_data: Sign data or files using a GPG/PGP private key stored in Pincer's vault. (Keyless Execution — agent never sees the key)
gpg_decrypt: Decrypt PGP-encrypted data using a vault-stored private key.

🔑 GPG Key Management

bash

# Generate a new GPG keypair (private key stored in vault)
pincer key generate --name "Release Signing" --email dev@example.com

# Import an existing PGP private key
pincer key import ./my-key.asc --passphrase "my-passphrase"

# List all stored GPG keys
pincer key list

# Export public key (safe to share)
pincer key export <key-id>

# Authorize an agent for signing
pincer agent authorize mybot gpg_sign_data --key <key-id>

(More callers coming soon!)

🚀 Quick Start

Prerequisites

Node.js 18+
macOS, Windows, or Linux with native keychain support

Installation

Option 1: Global Installation (Recommended)

bash

npm install -g pincer-mcp
# Now 'pincer' command is available system-wide

Option 2: Local Development

bash

git clone https://github.com/VouchlyAI/Pincer-MCP.git
cd Pincer-MCP
npm install
npm run build
npm link  # Makes 'pincer' command available locally

Setup Vault

bash

# 1. Initialize vault (creates master key in OS keychain)
pincer init

# 2. Store your real API keys (encrypted)
pincer set gemini_api_key "AIzaSyDpxPq..."
pincer set openai_api_key "sk-proj-..."

# 3. Register an agent and generate proxy token
pincer agent add openclaw
# Output: 🎫 Proxy Token: pxr_V1StGXR8_Z5jdHi6B-myT

# 4. Authorize the agent for specific tools
pincer agent authorize openclaw gemini_generate

Multi-Key Support

Store multiple keys for the same tool and assign them to different agents:

bash

# Store two different Gemini API keys
pincer set gemini_api_key "AIzaSy_KEY_FOR_CLAWDBOT..." --label key1
pincer set gemini_api_key "AIzaSy_KEY_FOR_MYBOT..." --label key2

# View all stored keys
pincer list

# Assign specific keys to each agent
pincer agent add clawdbot
pincer agent authorize clawdbot gemini_generate --key key1

pincer agent add mybot  
pincer agent authorize mybot gemini_generate --key key2

# View agent permissions
pincer agent list

Result: clawdbot uses key1, mybot uses key2 - perfect for rate limiting or cost tracking!

Run the Server

bash

npm run dev

Configure Your Agent

Give your agent the proxy token (not the real API key):

bash

export PINCER_PROXY_TOKEN="pxr_V1StGXR8_Z5jdHi6B-myT"

Tool-to-Secret Name Mappings

When storing secrets, you must use the correct secret name for each tool. See the Tool Mappings Guide for a complete reference.

When you run pincer agent authorize myagent gemini_generate, Pincer will inject the gemini_api_key secret when that tool is called.

Make a Tool Call

Your agent sends requests with the proxy token in the body:

json

{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "name": "gemini_generate",
    "arguments": {
      "prompt": "Hello world",
      "model": "gemini-2.0-flash"
    },
    "_meta": {
      "pincer_token": "pxr_V1StGXR8_Z5jdHi6B-myT"
    }
  }
}

Pincer maps the proxy token to the real API key and executes the call securely.

🏗️ Architecture

Two-Tiered Vault System

Tier 1: Master Key (OS Keychain)

Stored in macOS Keychain, Windows Credential Manager, or GNOME Keyring
Never touches the filesystem
Accessed only for encryption/decryption

Tier 2: Encrypted Store (SQLite)

Database at ~/.pincer/vault.db
Three tables:
- secrets: Real API keys (AES-256-GCM encrypted)
- proxy_tokens: Proxy token → Agent ID mappings
- agent_mappings: Agent ID → Tool authorization

Authentication Flow

code

Request (_meta.pincer_token: pxr_xxx)
  ↓
Gatekeeper: Extract proxy token from body
  ↓
Vault: Resolve pxr_xxx → agent_id → tool_name → real_api_key
  ↓
Injector: JIT decrypt & inject real key
  ↓
Caller: Execute external API call
  ↓
Scrubber: Overwrite key in memory with zeros
  ↓
Audit: Log to tamper-evident chain

🔐 Security & Compliance

Pincer is built for enterprise-grade security:

Hardware-Backed Cryptography: Master encryption keys never leave the OS-native keychain.
Proxy Token Isolation: Agents only handle ephemeral pxr_ tokens; they never touch real credentials.
JIT Decryption: Secrets are decrypted only for the duration of the API call.
Zero-Footprint Memory: Sensitive data is scrubbed (zeroed out) from memory immediately after use.
Fine-Grained Authorization: Strict per-agent, per-tool access control policies.
Tamper-Evident Audit Log: Append-only tool call history with SHA-256 chain-hashing.
Hardened Execution: Schema validation on all inputs and protected environment execution.
Stdio Compatible: Fully compatible with the standard Model Context Protocol transport.

🔍 Audit Logs

Every tool call is logged to ~/.pincer/audit.jsonl with both UTC and Local timestamps, plus character counts and estimated token usage:

json

{
  "agentId": "openclaw",
  "tool": "gemini_generate",
  "duration": 234,
  "status": "success",
  "input_chars": 156,
  "output_chars": 423,
  "estimated_input_tokens": 39,
  "estimated_output_tokens": 106,
  "timestamp_utc": "2026-02-05T08:32:00.000Z",
  "timestamp_local": "2/5/2026, 2:02:45 PM",
  "chainHash": "a1b2c3d4e5f6g7h8",
  "prevHash": "0000000000000000"
}

Token Estimation: Pincer automatically estimates token usage using a 4:1 character-to-token ratio (~4 characters per token average). This provides consistent cost tracking across all AI providers without relying on provider-specific APIs.

code


Chain hashes provide tamper detection - any modification breaks the SHA-256 chain.

## 🧪 Development

```bash
# Install dependencies
npm install

# Run tests
npm test

# Run with watch mode
npm run dev

# Build for production
npm run build

📚 Documentation

Setup Guide - Getting started with Pincer-MCP
IDE Integration - Use Pincer with VSCode, Claude Desktop, Cursor, and more
OpenClaw Integration - Integrate Pincer with OpenClaw agents
Testing Guide - Comprehensive test suite documentation
Capabilities Reference - Full API and feature documentation
Security Policy - Vulnerability reporting and security best practices
CHANGELOG - Version history and release notes

🤝 Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

📄 License

BSL 1.1 (Business Source License) — See LICENSE for details. Converts to Apache 2.0 on 2028-04-01.

Model Context Protocol - The standard for AI tool integration.
keytar - Secure cross-platform keychain access.
better-sqlite3 - High-performance local persistence.

Built with ❤️ for a more secure AI future.

Install

Configuration

Environment variables

PINCER_PROXY_TOKENrequiredsecret

Proxy token for agent authentication (generated via 'pincer agent add')

OPENWEBUI_URL

Optional: Base URL for OpenWebUI instance

OPENAI_COMPATIBLE_URL

Optional: Base URL for OpenAI-compatible instance

AUDIT_LOG_PATH

Optional: Custom path for audit logs (default: ~/.pincer/audit.jsonl)

claude_desktop_config.json

{
  "mcpServers": {
    "pincer": {
      "command": "npx",
      "args": [
        "-y",
        "pincer-mcp@0.1.5"
      ],
      "env": {
        "PINCER_PROXY_TOKEN": "<YOUR_PINCER_PROXY_TOKEN>",
        "OPENWEBUI_URL": "<YOUR_OPENWEBUI_URL>",
        "OPENAI_COMPATIBLE_URL": "<YOUR_OPENAI_COMPATIBLE_URL>",
        "AUDIT_LOG_PATH": "<YOUR_AUDIT_LOG_PATH>"
      }
    }
  }
}