Add screenshots to any AI agent
Grabbit is a hosted MCP server and a one-line CLI. Point your agent at a URL and get back a clean, hosted image, including the cookie walls and JavaScript-heavy pages a local headless browser chokes on. No Chromium to install.
Two ways to connect
Use the hosted MCP server for agent frameworks, or the CLI for shell agents and scripts. Both talk to the same flat-priced capture engine.
# Hosted MCP server (Streamable HTTP), OAuth or Bearer key
claude mcp add --transport http grabbit https://mcp.grabbit.live
# Endpoint: https://mcp.grabbit.live
# Tools: grab · get_grab · list_grabs · get_usage · get_pricing
# Auth: OAuth 2.1 (claude.ai connectors work out of the box) or sk_ key# One-line CLI (STDIO/local), great for shell agents and scripts
npx grabbit.live login # browser sign-in, keys saved locally
npx grabbit.live nyt.com # → ./nyt.com.png + a hosted URL
npx grabbit.live mcp # print MCP config for your agentPick your agent
Claude Code
Add the hosted MCP server with one command (OAuth, no key). Claude verifies the UI it just built, then fixes it.
Read the guide →GuideOpenAI Codex
One config.toml block points Codex at the hosted endpoint. No local process, no browser to install.
Read the guide →GuideOpenClaw
It already browses. Drop in the Grabbit skill so it gets the captures its own browser is blocked on.
Read the guide →GuideCursor
One mcp.json entry, OAuth so no key in the file. Commit it to the repo and the whole team gets it.
Read the guide →Using something else? Any MCP client (including the claude.ai desktop and web connectors) can add the hosted server, and everything else can call the REST API or the CLI.
Captures your agent's own browser can't get
Most agents can already drive a headless browser. It works on simple pages and quietly fails on the ones that matter: cookie and consent walls, bot detection, login and paywall gates, and heavy client-rendered pages. Grabbit runs the capture on hosted infrastructure, so the shot comes back clean, and the agent needs no browser of its own.
Local headless browser
- Needs Chromium installed in the agent's runtime
- Blank or blocked on bot-walled and consent-gated pages
- You own the timeouts, retries, and scaling
- Breaks in CI, sandboxes, and remote runners with no browser
Grabbit
- One call, nothing to install or run
- Dismisses cookie and consent banners automatically
- Handles lazy-loaded, JavaScript-heavy, and full-page captures
- Gets many sites a local headless browser is blocked on
Flat pricing for agents that loop
When an agent captures a page on every step, per-call metering becomes anxiety. Grabbit is one flat rate: $50 per year for 25,000 captures, with prepaid packs that never reset. No monthly quota to forfeit, no usage cliff.
Predictable to call and to handle
Bearer or OAuth auth, a predictable JSON response, an inline image, and a structured error envelope make success and failure easy for an agent to reason about. The grab result even returns remaining credits, so the agent can top up before it runs dry.
Screenshots for agents FAQ
- What is the best screenshot tool for an AI agent?
- The right one fits how your agent runs. If it can drive a local browser and only hits simple pages, a Playwright or Puppeteer MCP server works. If you want zero local setup, captures that survive cookie walls and JS-heavy pages, and flat pricing while the agent loops, a hosted tool like Grabbit fits better. It adds no Chromium to your runtime and returns a hosted image in one call.
- How do I add a screenshot tool to my agent via MCP?
- Point your MCP client at Grabbit's hosted server: https://mcp.grabbit.live over Streamable HTTP. In Claude Code that is one `claude mcp add` command; in Codex it is a short config.toml block. Authenticate with OAuth or a Bearer key. The server exposes a grab tool plus get_grab, list_grabs, get_usage, and get_pricing.
- Does my agent need a local headless browser?
- No. Grabbit runs the capture on hosted infrastructure, so there is no Chromium, Playwright, or Puppeteer to install. That also means it works in CI, sandboxes, and remote runners where a local browser has nothing to drive.
- Which agents and clients are supported?
- Any MCP client, including Claude Code, the claude.ai desktop and web connectors, Cursor, and OpenAI Codex, plus open agents like OpenClaw via a skill. Anything else can call the REST API or the one-line CLI.
- Can an agent screenshot any website?
- It captures public URLs and handles cookie and consent banners, lazy-loaded and JavaScript-heavy pages, and many sites that block a local headless browser. Private and internal addresses (including localhost) are blocked by SSRF protection, and a few hard targets behind aggressive IP blocking still resist. No service captures literally everything.
- What does it cost?
- A flat $0.002 per live capture on the $50/year plan (25,000 captures), with prepaid packs that never reset or expire. Test-environment captures are free, so an agent can be wired up before a card is added.
Give your agent eyes in two minutes
Free test key, one install, and any agent can screenshot the web.