Drive Coasty from Cursor and Claude with the MCP Server
Most agentic workflows today are confined to APIs. You send JSON and get JSON, which is reliable but lacks the ability to interact with real UIs. The Coasty Computer Use API solves this by letting an agent see a screen and perform human-like actions. The new MCP server bridges that capability to Cursor, Claude Desktop, and other Model Context Protocol clients, so you can build agents that drive desktops, browsers, and terminals directly from AI editors.
How the MCP server works
The MCP server exposes Coasty endpoints as tools. When you ask Cursor or Claude to 'open https://coasty.ai/docs in a browser' the local MCP server forwards a request to Coasty's cloud machines. Coasty provisions a cloud VM, starts a session, and runs a computer use agent that captures the screen, predicts actions, and reports results back to the MCP server. The editor receives structured tool outputs and builds its own reasoning before planning the next action. This keeps all sensitive tokens and API keys local while leveraging Coasty's vision-based reasoning.
# Install the MCP server from GitHub (example) and configure your editor
# For Cursor, add the MCP server config to ~/.cursor/mcp_config.json:
# {
# "mcpServers": {
# "coasty": {
# "command": "npx",
# "args": ["-y", "@coasty/mcp-server", "--api-key", "$COASTY_API_KEY"],
# "env": {
# "COASTY_API_KEY": "$(cat ~/.coasty_key)"
# }
# }
# }
# }
# Ensure you set COASTY_API_KEY from the environment:
# export COASTY_API_KEY="your_key_here"
# Then in Cursor, ask: 'Open https://coasty.ai/docs in the browser and take a screenshot.'Key MCP client features
- ●Read COASTY_API_KEY from your local environment; never hardcode it in configs.
- ●The MCP server runs on your machine, so API keys stay local and secrets stay private.
- ●All agent steps on Coasty are billed $0.05 per step via the Coasty prepaid wallet.
- ●You can chain multiple tools: start a session, capture, predict, act, and repeat until done.
- ●Use conditional logic and error handling in your prompts to control retries and timeouts.
With the MCP server, Coasty endpoints become tools in your AI editor. Ask in natural language, get structured results, and let the agent drive real UIs.
Where this beats brittle automation
Traditional automation relies on brittle selectors like IDs or XPath. When UIs change, these break. Coasty's computer use agent sees the screen like a human, understands context, and adapts to layout shifts. It can click buttons, fill forms, navigate frames, and even handle dynamic content. This makes it far more resilient for complex workflows that span desktop apps, browsers, and terminals. By combining Coasty's vision-based actions with Cursor or Claude's reasoning, you get agents that can both understand what to do and execute it on real interfaces.
Now you can build agents that act directly on desktops and browsers from Cursor, Claude Desktop, and other MCP clients. Start by installing the MCP server, setting COASTY_API_KEY from your environment, and asking your editor to drive a browser or app. Explore workflows and task runs on Coasty to scale your automation. Get a key at https://coasty.ai/developers.