MCP server
screenshotbase ships a hosted Model Context Protocol (MCP) server, so AI agents and assistants can call the API as native tools — no SDK or glue code required.
https://api.screenshotbase.com/mcp
The endpoint speaks the streamable HTTP transport. All requests — including listing the available tools — require your API key, sent as the apikey header. You can get a free API key here.
MCP clients that support MCP authorization (OAuth 2.1) can sign in instead of sending a key: see Sign in instead of an API key.
Connect
Using Claude Code:
claude mcp add --transport http screenshotbase https://api.screenshotbase.com/mcp --header "apikey: YOUR_API_KEY"
Or add the server to any MCP-capable client (Claude Desktop, Cursor, VS Code, ...):
{
"mcpServers": {
"screenshotbase": {
"url": "https://api.screenshotbase.com/mcp",
"headers": { "apikey": "YOUR_API_KEY" }
}
}
}
Sign in instead of an API key
MCP clients that support MCP authorization (OAuth 2.1) can connect without pasting a key, for example a custom connector in Claude (Settings → Connectors → Add custom connector, URL https://api.screenshotbase.com/mcp). Without a key, the endpoint answers 401 with a WWW-Authenticate header that names the protected resource metadata, https://api.screenshotbase.com/.well-known/oauth-protected-resource/mcp. The authorization server is https://screenshotbase.com (metadata at https://screenshotbase.com/.well-known/oauth-authorization-server, endpoints under https://app.screenshotbase.com/oauth/); clients register dynamically or with a client ID metadata document, and PKCE (S256) is required.
You sign in on app.screenshotbase.com and approve the connection, as the team's owner or an admin. The access token is a key of your team for 60 minutes, renewed with the refresh token, so tool calls count against your plan like any other request. Disconnect an app at any time under API keys → Connected apps in the dashboard. For your own code, use an API key.
Available tools
The tools are generated from the same OpenAPI specification that describes the REST API, so they always match the documented endpoints, parameters and responses.
| Tool | Endpoint | Description |
|---|---|---|
take | GET /v1/take | Take a screenshot, PDF or markdown of a URL |
takePost | POST /v1/take | Take a screenshot, PDF or markdown of HTML or a URL |
getStatus | GET /v1/status | Account quota status |
listDevices | GET /v1/devices | List the devices of viewport_device |
listJobs | GET /v1/jobs | List your async jobs |
getJob | GET /v1/jobs/{job_id} | Get an async job |
getJobResult | GET /v1/jobs/{job_id}/result | Download an async job's result |
deleteJob | DELETE /v1/jobs/{job_id} | Cancel and delete an async job |
Screenshots, PDFs and markdown
With format: "markdown", the tools return the page's content as markdown text, which agents can use directly (see URL to markdown).
A screenshot arrives as image content and a PDF as an embedded resource (screenshotbase://<tool>/<file name>, for example screenshotbase://take/example.pdf, after a text that names it), so the agent sees the file itself. A tool result holds files of up to 3 MB and images of up to 8,000 px on each side; a full-page screenshot gets there quickly. A larger result answers with an error that asks for upload: true, and is not billed.
With upload: true, the tool returns a JSON object with the public url of the stored file instead, which the agent can fetch, show or pass on. Every render costs one request with or without upload. Stored files are public to anyone with the URL.
Time limit
A tool call must finish within about 12 seconds, because MCP clients wait less for a tool than HTTP clients for a response: delay is at most 7 and timeout at most 60 there. For slower pages, pass async: true and fetch the result with getJob and getJobResult (see Async Jobs & Webhooks), or call the REST API, which waits up to 110 seconds.
Quotas and errors
Tool calls are metered exactly like REST requests: they consume your plan quota and return the same status codes and error responses (401, 422, 429, ...). If a call fails, the tool result contains the API's error message including hints on how to proceed.