Skip to content
GFlow4AI
VIGet for Windows
Guide/Connect AI agents

Connect an AI agent

5 steps · 3 screenshotsUpdated Try the web sample
On this page · 5 steps

Choose a local client, enable the server, and verify read-only access before generation.

Before you start

  • Run GFlow4AI (the app shows Flow Studio) on Windows; the web preview opens no real port.
  • Check that the client runs locally, not in a cloud workspace.

Steps

  1. Check where the agent runs

    The server binds only IPv4 127.0.0.1, not LAN or Internet. Claude cloud, Lovable, Manus and ChatGPT cannot call your Windows loopback; use manual prompt handoff instead. WSL, containers, SSH workspaces and hosted n8n may also have a different loopback.

  2. Enable the server

    Open Settings → API & agent, then turn on Enable API server. Check Running · this computer only and copy the displayed address; the default is http://127.0.0.1:47831. Keep the desktop app running.

    Settings › API & agentReal app screen · web sample data · private details blurred

    Settings › API & agent · Real app screen · web sample data · private details blurred

  3. Create a separate key

    Create one named key per client with only the needed scopes and a daily cap. Save its full secret before closing the reveal dialog; see API keys. Do not use Google passwords or cookies.

    Creating an API key with its limitsReal app screen · web sample data · private details blurred

    Creating an API key with its limits · Real app screen · web sample data · private details blurred

  4. Choose MCP or REST

    Use MCP setup for a local client that supports authenticated HTTP, or REST for scripts. Both use this exact header, with the full local key replacing the placeholder. Query tokens and X-API-Key do not authenticate.

    http
    Authorization: Bearer fsk_…
    Ready-to-copy MCP setup for AI agentsReal app screen · web sample data · private details blurred

    Ready-to-copy MCP setup for AI agents · Real app screen · web sample data · private details blurred

  5. Test reading before generation

    Call flow_status or GET /v1/status, then inspect models. Generation may cost credits: keep the same idempotency identifier on retries and follow the existing job. Review agent limits before allowing unattended work.

When it worked

Your local client can read authenticated status without exposing the server to the network.

If something goes wrong

Why can a cloud connector not connect?

Its 127.0.0.1 is not your PC. No public HTTPS, OAuth bridge or tunnel setup is implemented here.

Why does the web preview not respond?

It simulates settings without opening a socket. Start the Windows app and enable its server.

Why is the connection refused?

Check the running status and current port, then use 127.0.0.1 explicitly. Update every client after changing the port.