Kernelreview, pricing and limits

For teams building browser-using agents, Kernel replaces a self-hosted Playwright fleet with a managed MCP connector: register it once in Claude Code, Cursor, or Windsurf and every session gets stealth proxies, CAPTCHA solving, and persistent auth automatically. Sessions run up to 72 hours, and 1,000+ companies including Cash App and Rye already route agent browsing through it.

Kernel is a hosted Model Context Protocol server that lets Claude, Cursor, VS Code, and other MCP-compatible agents control cloud Chromium browsers through 18 tools covering sessions, profiles, proxies, and Playwright execution. Built on unikernels, it delivers sub-150ms cold starts and connects via Streamable HTTP with OAuth 2.0 authentication.

Maker: Kernel · Protocol: MCP · Auth: oauth

Compatible agents: Claude Code, Claude Desktop, Cursor, VS Code (via MCP), Windsurf, Zed, Goose, Playwright / Puppeteer scripts, Browser Use, Stagehand, Magnitude, Notte agent frameworks, Anthropic / OpenAI / Gemini computer-use loops

Required runtime: Any MCP-compatible client (Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Zed, Goose), Kernel CLI for one-line install: brew install onkernel/tap/kernel, Node.js 18+ if self-hosting the mcp-remote stdio bridge

About Kernel

Kernel is browser infrastructure for AI agents, built by Catherine Jue (CEO) and Rafael Garcia (CTO) and backed by Accel and Y Combinator's Summer 2025 batch. Unlike a browsing tool a person opens directly, Kernel is consumed programmatically: agent frameworks and MCP clients call its API or its hosted Model Context Protocol server to spin up and control real Chromium sessions. The company says more than 1,000 organizations, including Cash App and Rye, already route agent browsing through it. The platform runs sandboxed Chromium browsers on a unikernel architecture rather than standard containers, which the company and third-party benchmarks credit for unusually fast cold starts against container-based competitors like Browserbase. Every session can carry datacenter, ISP, residential, or mobile proxies with automatic CAPTCHA solving, and idle sessions in Standby Mode or a warmed pool incur no usage charges. The hosted MCP server at mcp.onkernel.com exposes a wide set of model-facing tools, covering browser and pool management, persistent Profiles, Managed Auth with TOTP-backed credentials, Playwright/TypeScript execution, shell access inside the browser VM, and documentation search, authenticated over OAuth 2.0/2.1 via Clerk. Any MCP-compatible client can connect, including Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Zed, and Goose, and the underlying API is built to work with Playwright, Puppeteer, Browser Use, Stagehand, Magnitude, Notte, and the Anthropic, OpenAI, and Gemini computer-use loops. Typical uses include letting an MCP agent click and read live pages instead of relying on stale training data, running CAPTCHA-gated scraping jobs unattended, and persisting a logged-in session across scheduled agent runs. Kernel is usage-based, with a free entry tier plus paid Hobbyist, Startup, and custom Enterprise plans that scale by included monthly credit; usage beyond that credit bills per second by browser type, and idle time and proxies are never charged (see the cost FAQ below for exact figures). The MCP server and the kernel-images browser build are open source under the MIT license on GitHub, while the managed platform itself remains closed-source. The team ships on a roughly weekly cadence: recent changelog entries added OAuth/PKCE improvements and private browser networking to the MCP server, launched a Terraform provider, and added a get_telemetry action for archived session data, all logged in Kernel's public changelog through August 2026.

Key Features

• Sub-150ms unikernel cold starts: Browsers run on a unikernel architecture instead of standard containers, landing cold starts around 150ms, about 2.6x faster than container-based competitors in third-party benchmarks.
• Standby Mode with zero idle cost: Idle browsers in a pool automatically suspend and resume on the next request, so idle time is never billed.
• Built-in stealth and CAPTCHA solving: Datacenter, ISP, residential, and mobile proxies pair with automatic CAPTCHA solving and improved fingerprint coherence.
• Persistent Profiles: Saved browser profiles carry cookies, logins, and extensions across sessions and separate agent invocations.
• Managed Auth with TOTP: Agents can log into a site through a hosted auth flow without ever seeing the raw credential, with TOTP code retrieval for 2FA-gated accounts.
• Full session observability: Live session view, MP4 replay recordings with chapter markers, and OTLP-exportable telemetry cover console output, network activity, and user interactions.
• SOC 2 compliant, HIPAA-ready: The platform carries SOC 2 compliance and is built to HIPAA-ready standards for regulated automation workloads.
• Broad framework compatibility: Integrates with the major browser-automation libraries and computer-use agent loops developers already build with.

Use Cases

• Give an MCP agent hands on the live web: Register the hosted MCP endpoint in Claude Code or Cursor so the agent can call manage_browsers and computer_action directly to click, type, and read real pages instead of relying on stale training data.
• Run CAPTCHA-gated scraping jobs unattended: Route requests through Kernel's residential and ISP proxies with automatic CAPTCHA solving so a scraping agent keeps working on sites that would otherwise block headless traffic.
• Persist a logged-in session across agent runs: Create a browser Profile once so cookies and auth survive between invocations, letting a scheduled agent skip re-login on every run.
• Debug a failed automation from a video replay: Pull the MP4 replay for a session via manage_replays to see exactly where a Playwright script or computer-use loop diverged from the expected page state.

Install

brew install onkernel/tap/kernel && kernel mcp install --target claude-code

Requirements

• Install the Kernel CLI (brew install onkernel/tap/kernel), or skip it and point any MCP client at the hosted endpoint via npx -y mcp-remote https://mcp.onkernel.com/mcp
• Create a Kernel account/organization at kernel.sh; the MCP server authenticates via OAuth 2.0/2.1 through Clerk
• An MCP-compatible client: Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Zed, or Goose
• Docker or an Unikraft unikernel runtime only if self-hosting the open-source kernel-images browser build

Actions

manage_browsers

Create, update, list, get, and delete cloud browser sessions, and read archived telemetry for a session.

{"tool": "manage_browsers", "arguments": {"action": "create", "headless": false, "stealth": true, "persistence": {"id": "my-profile"}}}

• action (string) — required: One of create, update, list, get, delete, get_telemetry.
• headless (boolean): Run without a visible display; headful sessions cost more but support computer_action.
• stealth (boolean): Enable anti-detection fingerprint coherence for the session.
• profile_id (string): Attach a saved Profile so cookies and logins persist.
• proxy_id (string): Attach a configured proxy (datacenter, ISP, residential, or mobile) to the session.
• session_id (string): Required for update, get, delete, and get_telemetry actions.

manage_profiles

Set up a persistent browser profile via a guided live session, and search, list, get, or delete existing profiles.

{"tool": "manage_profiles", "arguments": {"action": "setup", "name": "github-login"}}

• action (string) — required: One of setup, list, search, get, delete.
• name (string): Human-readable name for the profile.
• profile_id (string): Required for get and delete actions.
• query (string): Search term used with the search action.

manage_projects

Create, list, get, update, and delete organization projects, and inspect their resource limits.

{"tool": "manage_projects", "arguments": {"action": "list"}}

• action (string) — required: One of create, list, get, update, delete.
• project_id (string): Required for get, update, and delete actions.
• name (string): Project display name, used with create/update.

manage_api_keys

Create, list, get, update, and delete organization-wide or project-scoped API keys. The plaintext key is returned once, at creation.

{"tool": "manage_api_keys", "arguments": {"action": "create", "scope": "project", "project_id": "proj_123"}}

• action (string) — required: One of create, list, get, update, delete.
• scope (string): org or project.
• project_id (string): Required when scope is project.
• key_id (string): Required for get, update, and delete actions.

manage_browser_pools

Create, list, get, delete, and flush pools of pre-warmed browsers, and acquire or release a browser from a pool.

{"tool": "manage_browser_pools", "arguments": {"action": "acquire", "pool_id": "pool_abc"}}

• action (string) — required: One of create, list, get, delete, flush, acquire, release.
• pool_id (string): Required for get, delete, flush, acquire, release.
• max_concurrent_sessions (number): Unified concurrency cap for the pool.

manage_proxies

Create, list, get, check, and delete proxy configurations across datacenter, ISP, residential, mobile, and custom proxy types.

{"tool": "manage_proxies", "arguments": {"action": "create", "type": "residential"}}

• action (string) — required: One of create, list, get, check, delete.
• type (string): datacenter, isp, residential, mobile, or custom.
• proxy_id (string): Required for get, check, and delete actions.

manage_replays

Start, stop, and list MP4 video replay recordings for a browser session. Requires a paid Kernel plan.

{"tool": "manage_replays", "arguments": {"action": "start", "session_id": "sess_123"}}

• action (string) — required: One of start, stop, list.
• session_id (string) — required: The browser session to record.

manage_extensions

List and delete browser extensions uploaded to the organization.

{"tool": "manage_extensions", "arguments": {"action": "list"}}

• action (string) — required: One of list, delete.
• extension_id (string): Required for the delete action.

manage_apps

List and search deployed apps, invoke app actions, and manage and inspect deployments and invocation results.

{"tool": "manage_apps", "arguments": {"action": "invoke", "app_id": "app_123", "action_name": "run"}}

• action (string) — required: One of list, search, invoke, get_deployment, list_deployments, delete_deployment, get_invocation.
• app_id (string): Target app for invoke and deployment actions.
• action_name (string): The specific app action to invoke.

manage_auth_connections

Create, list, get, and delete managed-auth connections, and drive the login/submit/wait flow that signs an agent into a site without exposing credentials.

{"tool": "manage_auth_connections", "arguments": {"action": "login", "domain": "example.com"}}

• action (string) — required: One of create, list, get, delete, login, submit, wait.
• domain (string): Domain to filter or target for the auth connection.
• connection_id (string): Required for get, delete, submit, and wait actions.

manage_credentials

Create, list, get, update, and delete stored credentials, and fetch current TOTP codes for a stored credential.

{"tool": "manage_credentials", "arguments": {"action": "get_totp", "credential_id": "cred_123"}}

• action (string) — required: One of create, list, get, update, delete, get_totp.
• credential_id (string): Required for get, update, delete, and get_totp actions.

manage_credential_providers

Create, list, get, update, and delete external credential providers, such as a connected 1Password vault.

{"tool": "manage_credential_providers", "arguments": {"action": "list"}}

• action (string) — required: One of create, list, get, update, delete.
• provider_id (string): Required for get, update, and delete actions.

get_connection_context

Inspect the authenticated principal, organization, credential scope, and effective project scope for the current MCP session.

{"tool": "get_connection_context", "arguments": {}}

computer_action

Send mouse, keyboard, clipboard, and screenshot controls to a live browser session, for click, type, press_key, scroll, move, get_position, read_clipboard, write_clipboard, and screenshot operations.

{"tool": "computer_action", "arguments": {"session_id": "sess_123", "operation": "click", "x": 420, "y": 180}}

• session_id (string) — required: The browser session to control.
• operation (string) — required: One of click, type, press_key, scroll, move, get_position, read_clipboard, write_clipboard, screenshot.
• x (number): X coordinate for click, move, or scroll.
• y (number): Y coordinate for click, move, or scroll.
• text (string): Text to type or write to clipboard.

browser_curl

Send an HTTP request through a live browser session's own Chrome network stack, preserving its cookies, proxy, and fingerprint.

{"tool": "browser_curl", "arguments": {"session_id": "sess_123", "url": "https://example.com/api", "method": "GET"}}

• session_id (string) — required: The browser session whose network stack issues the request.
• url (string) — required: Target URL.
• method (string): HTTP method.

execute_playwright_code

Run Playwright/TypeScript code against an existing browser session and return the result.

{"tool": "execute_playwright_code", "arguments": {"session_id": "sess_123", "code": "return await page.title();"}}

• session_id (string) — required: The browser session to execute against.
• code (string) — required: Playwright/TypeScript code, executed with `page` in scope.

exec_command

Run a shell command inside the browser session's VM and return decoded stdout/stderr.

{"tool": "exec_command", "arguments": {"session_id": "sess_123", "command": "ls /tmp"}}

• session_id (string) — required: The browser session's VM to run the command in.
• command (string) — required: Shell command to execute.

search_docs

Search Kernel's own documentation from within an MCP client.

{"tool": "search_docs", "arguments": {"query": "standby mode pricing"}}

• query (string) — required: Search terms.

open_auth_login

Open a secure interactive Managed Auth MCP App after user consent, for clients that declare MCP Apps support.

{"tool": "open_auth_login", "arguments": {"connection_id": "conn_123"}}

• connection_id (string) — required: The managed-auth connection to open the interactive login for.

How to Invoke

Agents connect to the hosted MCP endpoint (https://mcp.onkernel.com/mcp) over Streamable HTTP, or via stdio through the mcp-remote fallback, and call the 18 exposed tools (manage_browsers, computer_action, execute_playwright_code, etc.) by name once the server is registered in the client's MCP config.

Pricing

Developer tier is free with $5/month of included usage credit. Hobbyist is $30/month with $10 of credit, Startup is $200/month with $50 of credit, and Enterprise is custom. Beyond credits, usage bills per second: headless ~$0.06/hour, headful ~$0.48/hour, headful+GPU ~$2.88/hour. Idle time (Standby Mode) and proxies are never charged.

Strengths

• Idle time is never billed: Standby Mode and pooled browsers only incur usage charges while a session is actively working.
• A single hosted MCP endpoint bundles the whole tool set, so an agent gets browser control, Playwright execution, and shell access without stitching together separate integrations.
• Ships on a roughly weekly cadence: the public changelog logged ten dated feature or fix entries between late May and mid-August 2026.
• SOC 2 compliance and HIPAA-readiness make it usable for regulated automation workloads that would otherwise need a self-hosted setup.

Weaknesses

• The core managed platform is closed-source; only the kernel-images browser build and the MCP server itself are open source under MIT.
• Session replay recording requires a paid plan, and the free Developer tier keeps only 1-day retention on other data.
• Headful-with-GPU sessions cost roughly 48x more per second than headless (about $0.0008 vs $0.0000167 per second), so GPU-heavy workloads need active usage monitoring.
• Pricing is usage-based rather than flat, which makes cost harder to predict for spiky, high-concurrency workloads than a fixed-seat product.

View the official Kernel skill page