Kernel MCP Server for AI Tools & Automation

Overview

The Kernel MCP Server is an open-source Model Context Protocol (MCP) server that lets MCP-capable AI assistants and developer tools securely access Kernel platform resources and cloud browser automation. It acts as a bridge between AI agents (for example, Claude, Cursor, or other MCP clients) and Kernel’s managed services: app deployments, headless Chromium sessions, Playwright execution, documentation search, and telemetry.

You can use the hosted endpoint at https://mcp.onkernel.com/mcp without deploying anything yourself, or run the code from the repository on GitHub. Authentication is handled via OAuth 2.0 (Clerk) with dynamic client registration so each user authorizes access to their Kernel account before tools can interact with resources.

Features

• Streamable HTTP MCP endpoint designed for reliable agent connections
• stdio fallback via mcp-remote for clients that don’t support remote MCP transports
• Create, deploy, and manage Kernel apps and cloud deployments
• Launch and control headless Chromium browser sessions for web automation
• Execute arbitrary Playwright scripts against live browser sessions
• Automatic video recording of browser automation sessions (replay)
• Search Kernel documentation and inject contextual snippets into agent prompts
• Invocation and deployment monitoring / observability
• Open-source codebase — production instance is hosted, but you can self-host

Repository: https://github.com/onkernel/kernel-mcp-server

Installation / Configuration

Use the hosted endpoint (recommended) or deploy the server yourself from the GitHub repo.

Recommended transports:

• Streamable HTTP (preferred): https://mcp.onkernel.com/mcp
• stdio via mcp-remote (fallback): npx -y mcp-remote https://mcp.onkernel.com/mcp

Quick setup with Kernel CLI (recommended for developer workflows)

# macOS Homebrew
brew install onkernel/tap/kernel

# or npm
npm install -g @onkernel/cli

# Add Kernel MCP to a supported target
kernel mcp install --target <target>

If you prefer to use mcp-remote directly:

# Launch a local stdio bridge to the hosted server
npx -y mcp-remote https://mcp.onkernel.com/mcp

Authentication

• The server uses OAuth 2.0 via Clerk. When you connect from a client (e.g., Claude or Cursor) you will be redirected to authorize access to your Kernel resources.
• The Kernel CLI automates adding the MCP configuration to many tools’ config files and does not require manual JSON editing.

Supported install targets (via kernel CLI)

Available Tools / Resources

Agents connected to the MCP server can access Kernel-specific capabilities exposed as MCP tools/resources. Typical resources include:

• Deployments API: create, update, and rollback Kernel apps
• Browser sessions: spawn headless Chromium instances, navigate pages, interact with DOM
• Playwright runner: run custom Playwright code in an authenticated cloud browser
• Replays: recorded video of browser sessions for debugging and audits
• Documentation search: index and search Kernel docs for context injection
• Metrics & logs: monitor invocation counts and deployment health

The exact toolset presented to an agent depends on the MCP client’s UI and the permissions granted during OAuth authorization.

Use Cases

• Web automation by an AI assistant: have an agent open a headless Chromium session, log into a web app, and perform tasks such as submitting forms, scraping structured data, or validating flows. Sessions can be recorded for later review.
• End-to-end application operations: instruct an agent to deploy a Kernel app, run smoke tests in cloud browsers, and report deployment health and logs back into the chat context.
• Reproducible debugging: run a Playwright script via the MCP server to reproduce a bug against a live browser session and attach recorded video to the incident.
• Contextual augmentation for agents: search Kernel docs and dynamically inject relevant documentation snippets into the model context to improve answers or automation logic.
• CI and automation orchestration: integrate MCP-enabled agents into developer workflows to trigger deployments, run browser-driven test suites, and summarize results.

Connecting Common Clients (examples)

Claude (Pro/Max/Team/Enterprise)

1. Settings → Connectors → Add custom connector.
2. Name: Kernel, URL: https://mcp.onkernel.com/mcp
3. Click Connect and complete OAuth authorization.

Cursor

• Automatic install via Cursor’s MCP installer (deeplink) or manually add an MCP server configuration:

{
  "mcpServers": {
    "kernel": { "url": "https://mcp.onkernel.com/mcp" }
  }
}

OpenCode

• Add to config (~/.config/opencode/opencode.jsonc):

{
  "mcp": {
    "kernel": { "type": "remote", "url": "https://mcp.onkernel.com/mcp" }
  }
}

• Authenticate with the OpenCode CLI: opencode mcp auth kernel

Goose

• Use the provided deeplink or add a stdio extension that runs:

npx -y mcp-remote https://mcp.onkernel.com/mcp

Further information

• GitHub: https://github.com/onkernel/kernel-mcp-server
• Hosted endpoint: https://mcp.onkernel.com/mcp
• Blog / design notes: see Kernel blog for architecture rationale and implementation details.

If you want to self-host, the repository contains server code, Next.js routes, and example client integrations. Authentication and MCP transport configuration are documented in the repo README.