Remote MCP
VisFlow Agent Setup
Choose how to execute
Step 1 · Client
Choose your agent
Step 2 · Method
Choose a setup method
Step 3 · Usage
Choose how the agent should invoke VisFlow
Step 3 · Try it
Connect, then try a simple task
See the browser open a page to confirm your setup works.
Get your browser ready
Open your browser, sign in to VisFlow, and enable External Agent step control in the extension.
Check the connection
This checks the connection without changing any page.
Try it in your Agent
After setup, start a new conversation and send one of the tasks below.
Need help getting connected?
Tool missing: start a new conversation, then reopen your Agent if needed. Credential expired: sign in to VisFlow again, reopen this page from the extension, and update the same connection or personal plugin.
VisFlow opens, reads and interacts with pages. Your selected Agent decides which steps to take.
Quick troubleshooting
Delete only the existing entry named VisFlow and try again. Do not clear the entire MCP configuration or remove other extension servers.
Confirm that Codex CLI or Claude Code is installed and available on PATH. If unsure, use the graphical setup method on this page.
Restart the client and begin a new conversation. If the count remains 0, verify the HTTP transport, complete URL, Authorization value, and service availability.
The current configuration is not bound to this VisFlow extension device. Reopen this page from the extension and replace the old configuration with the newly generated complete URL.
The current sign-in credential may have expired. Sign in to VisFlow again, reopen this page from the extension, and replace the old entry with the newly generated configuration.
First confirm that the task begins with “Use VisFlow.” To avoid repeating this in later tasks, switch to Method B, save the current agent’s user-level global rule, restart, and begin a new conversation. Do not add MCP twice or install extra dependencies.
This is a short-poll status for the same task. The agent must continue querying the original commandId; it must not execute again, switch browsers, or ask whether it should keep waiting.