Skip to main content
One pre.dev MCP server, all the tools. Install it once and your agent gets browser_agent (browser automation) plus fast_spec / deep_spec (the Architect API) in one shot.

​
Quick Start

1

Have a pre.dev account

Setup uses browser OAuth — your first MCP call opens pre.dev to sign in and authorize. No API key needed for editor use.
2

Install the MCP server

Run this in your terminal:
Your browser opens to authorize — sign into pre.dev (if not already) and pick which account or organization to connect.
See the Claude Code MCP docs for MCP basics.
3

Try it

Ask your agent:
Your agent calls the browser_agent tool and returns structured JSON.

​
Example prompts

​
Live event streaming

Every step of a task emits an MCP notifications/message frame. The taskIndex field ties each event to a specific task in the batch. Your agent sees:
  • navigation — the agent loaded a URL
  • plan — the agent decided what to do next
  • action — a click, type, or scroll executed
  • screenshot — a screenshot was captured
  • done — the task completed
Event shapes are identical to the REST SSE stream — see Run a Task → SSE frames.

​
Auth

In editors, auth is browser OAuth — you sign in once and pick the account or org to connect. For programmatic MCP clients, an API key from pre.dev/projects/key also works as a predev-api-key header.

​
Troubleshooting

  • Restart your client after adding the MCP config
  • Verify the URL is exactly https://api.pre.dev/mcp
  • Check JSON syntax in your config file
  • Confirm your API key is valid at pre.dev/projects/key
  • Make sure the Authorization header starts with Bearer (with the space)
  • Solo accounts: use your personal key. Enterprise: use the org key from your org settings
  • Check your subscription is active
browser_task, browser_task_list, browser_task_get still work as aliases — old integrations continue working without changes.