Skip to main content

Open-source MCP security, aggregation, and monitoring. Single-user, self-hosted MCP proxy.

Project description

OpenEdison 🔒⚡️

The Secure MCP Control Panel

Connect AI to your data/software securely without risk of data exfiltration. Gain visibility, block threats, and get alerts on the data your agent is reading/writing.

OpenEdison solves the lethal trifecta problem, which can cause agent hijacking & data exfiltration by malicious actors.

Trifecta Security Risk Animation

📧 To get visibility, control and exfiltration blocker into AI's interaction with your company software, systems of record, DBs, Contact us to discuss.

Project Version Python Version License


Features ✨

  • 🛑 Data leak blocker - Edison automatically blocks any data leaks, even if your AI gets jailbroken
  • 🕰️ Deterministic execution - Deterministic execution. Guaranteed data exfiltration blocker.
  • 🗂️ Easily configurable - Easy to configure and manage your MCP servers
  • 📊 Visibility into agent interactions - Track and monitor your agents and their interactions with connected software/data via MCP calls
  • 🔗 Simple API - REST API for managing MCP servers and proxying requests
  • 🐳 Docker support - Run in a container for easy deployment

About Edison.watch 🏢

Edison helps you gain observability, control, and policy enforcement for all AI interactions with systems of records, existing company software and data. Prevent AI from causing data leakage, lightning-fast setup for cross-system governance.

Quick Start 🚀

The fastest way to get started:

# Installs uv (via Astral installer) and launches open-edison with uvx.
# Note: This does NOT install Node/npx. Install Node if you plan to use npx-based tools like mcp-remote.
curl -fsSL https://raw.githubusercontent.com/Edison-Watch/open-edison/main/curl_pipe_bash.sh | bash

Run locally with uvx: uvx open-edison That will run the setup wizard if necessary.

⬇️ Install Node.js/npm (optional for MCP tools)

If you need npx (for Node-based MCP tools like mcp-remote), install Node.js as well:

macOS

  • uv: curl -fsSL https://astral.sh/uv/install.sh | sh
  • Node/npx: brew install node

Linux

  • uv: curl -fsSL https://astral.sh/uv/install.sh | sh
  • Node/npx: sudo apt-get update && sudo apt-get install -y nodejs npm

Windows

  • uv: powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
  • Node/npx: winget install -e --id OpenJS.NodeJS

After installation, ensure that npx is available on PATH.

PyPI Install from PyPI

Prerequisites

  • Pipx/uvx
# Using uvx
uvx open-edison

# Using pipx
pipx install open-edison
open-edison

Run with a custom config directory:

open-edison run --config-dir ~/edison-config
# or via environment variable
OPEN_EDISON_CONFIG_DIR=~/edison-config open-edison run
Docker Run with Docker

There is a dockerfile for simple local setup.

# Single-line:
git clone https://github.com/Edison-Watch/open-edison.git && cd open-edison && make docker_run

# Or
# Clone repo
git clone https://github.com/Edison-Watch/open-edison.git
# Enter repo
cd open-edison
# Build and run
make docker_run

The MCP server will be available at http://localhost:3000 and the api + frontend at http://localhost:3001. 🌐

⚙️ Run from source
  1. Clone the repository:
git clone https://github.com/Edison-Watch/open-edison.git
cd open-edison
  1. Set up the project:
make setup
  1. Edit config.json to configure your MCP servers. See the full file: config.json, it looks like:
{
  "server": { "host": "0.0.0.0", "port": 3000, "api_key": "..." },
  "logging": { "level": "INFO", "database_path": "sessions.db" },
  "mcp_servers": [
    { "name": "filesystem", "command": "uvx", "args": ["mcp-server-filesystem", "/tmp"], "enabled": true },
    { "name": "github", "enabled": false, "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "..." } }
  ]
}
  1. Run the server:
make run
# or, from the installed package
open-edison run

The server will be available at http://localhost:3000. 🌐

🔌 MCP Connection

Connect any MCP client to Open Edison (requires Node.js/npm for npx):

npx -y mcp-remote http://localhost:3000/mcp/ --http-only --header "Authorization: Bearer your-api-key"

Or add to your MCP client config:

{
  "mcpServers": {
    "open-edison": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://localhost:3000/mcp/", "--http-only", "--header", "Authorization: Bearer your-api-key"]
    }
  }
}
🧭 Usage

API Endpoints

See API Reference for full API documentation.

🛠️ Development

Setup 🧰

Setup from source as above.

Run ▶️

Server doesn't have any auto-reload at the moment, so you'll need to run & ctrl-c this during development.

make run

Tests/code quality ✅

We expect make ci to return cleanly.

make ci
⚙️ Configuration (config.json)

Configuration ⚙️

The config.json file contains all configuration:

  • server.host - Server host (default: localhost)
  • server.port - Server port (default: 3000)
  • server.api_key - API key for authentication
  • logging.level - Log level (DEBUG, INFO, WARNING, ERROR)
  • mcp_servers - Array of MCP server configurations

Each MCP server configuration includes:

  • name - Unique name for the server
  • command - Command to run the MCP server
  • args - Arguments for the command
  • env - Environment variables (optional)
  • enabled - Whether to auto-start this server

🔐 How Edison prevents data leakages

🔱 The lethal trifecta, agent lifecycle management

Open Edison includes a comprehensive security monitoring system that tracks the "lethal trifecta" of AI agent risks, as described in Simon Willison's blog post:

The lethal trifecta diagram showing the three key AI agent security risks
  1. Private data access - Access to sensitive local files/data
  2. Untrusted content exposure - Exposure to external/web content
  3. External communication - Ability to write/send data externally
Privileged Access Management (PAM) example showing the lethal trifecta in action

The configuration allows you to classify these risks across tools, resources, and prompts using separate configuration files.

In addition to trifecta, we track Access Control Level (ACL) for each tool call, that is, each tool has an ACL level (one of PUBLIC, PRIVATE, or SECRET), and we track the highest ACL level for each session. If a write operation is attempted to a lower ACL level, it is blocked.

🧰 Tool Permissions (tool_permissions.json)

Defines security classifications for MCP tools. See full file: tool_permissions.json, it looks like:

{
  "_metadata": { "last_updated": "2025-08-07" },
  "builtin": {
    "get_security_status": { "enabled": true, "write_operation": false, "read_private_data": false, "read_untrusted_public_data": false, "acl": "PUBLIC" }
  },
  "filesystem": {
    "read_file": { "enabled": true, "write_operation": false, "read_private_data": true, "read_untrusted_public_data": false, "acl": "PRIVATE" },
    "write_file": { "enabled": true, "write_operation": true, "read_private_data": true, "read_untrusted_public_data": false, "acl": "PRIVATE" }
  }
}
📁 Resource Permissions (`resource_permissions.json`)

Resource Permissions (resource_permissions.json)

Defines security classifications for resource access patterns. See full file: resource_permissions.json, it looks like:

{
  "_metadata": { "last_updated": "2025-08-07" },
  "builtin": { "config://app": { "enabled": true, "write_operation": false, "read_private_data": false, "read_untrusted_public_data": false } }
}
💬 Prompt Permissions (`prompt_permissions.json`)

Prompt Permissions (prompt_permissions.json)

Defines security classifications for prompt types. See full file: prompt_permissions.json, it looks like:

{
  "_metadata": { "last_updated": "2025-08-07" },
  "builtin": { "summarize_text": { "enabled": true, "write_operation": false, "read_private_data": false, "read_untrusted_public_data": false } }
}

Wildcard Patterns ✨

All permission types support wildcard patterns:

  • Tools: server_name/* (e.g., filesystem/* matches all filesystem tools)
  • Resources: scheme:* (e.g., file:* matches all file resources)
  • Prompts: type:* (e.g., template:* matches all template prompts)

Security Monitoring 🕵️

All items must be explicitly configured - unknown tools/resources/prompts will be rejected for security.

Use the get_security_status tool to monitor your session's current risk level and see which capabilities have been accessed. When the lethal trifecta is achieved (all three risk flags set), further potentially dangerous operations are blocked.

Documentation 📚

📚 Complete documentation available in docs/

📄 License

GPL-3.0 License - see LICENSE for details.

Project details


Release history Release notifications | RSS feed

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

open_edison-0.1.76rc1.tar.gz (191.5 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

open_edison-0.1.76rc1-py3-none-any.whl (173.4 kB view details)

Uploaded Python 3

File details

Details for the file open_edison-0.1.76rc1.tar.gz.

File metadata

  • Download URL: open_edison-0.1.76rc1.tar.gz
  • Upload date:
  • Size: 191.5 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.1.0 CPython/3.12.3

File hashes

Hashes for open_edison-0.1.76rc1.tar.gz
Algorithm Hash digest
SHA256 9c8988b4fd8025583f8dd40508e1059510a1c469ba5a347f826a55c623f7e9ce
MD5 e44d51a5f454602ee6d69bdb4833227e
BLAKE2b-256 9e58b761cb5d8c7766b1c8d0b4f1b23cbac491505fd3fc83345bde5f69096081

See more details on using hashes here.

File details

Details for the file open_edison-0.1.76rc1-py3-none-any.whl.

File metadata

File hashes

Hashes for open_edison-0.1.76rc1-py3-none-any.whl
Algorithm Hash digest
SHA256 dcb2f6a5f97272bf057b14411f1f0ba488dd86b09f629c84a6315f896855db91
MD5 095ee9e7a6e6b8987fcefac8865b16ec
BLAKE2b-256 c299441c413f0f4cc65b5e7c27d29fe77c7b1cca3ef4c1fa352c487417c9c9be

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page