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

Set up local MCP

Updated 2026-11-04Try the web sample

Copy the app’s HTTP MCP template, check client compatibility, and discover the 14 tools.

Local MCP section
Local MCP section · Real app screen with web sample data · App labels are currently in Vietnamese

Before you start

  • Keep the Windows app and your agent on the same PC, with the server enabled and a saved API key.

Steps

  1. Use the supported transport

    The server provides MCP over Streamable HTTP at /mcp, with JSON responses and no server-initiated stream. It uses the same bearer key and scopes as REST. Copy the current address from Settings; the default endpoint is http://127.0.0.1:47831/mcp.

  2. Copy the exact app template

    In “Kết nối MCP”, copy the JSON template. The following is the app’s generated template at the default port. Replace fsk_… with your full key and adjust the port if changed. This is a shared template, not a verified native configuration for every client.

    {
  "mcpServers": {
    "flow-studio": {
      "type": "http",
      "url": "http://127.0.0.1:47831/mcp",
      "headers": {
        "Authorization": "Bearer fsk_…"
      }
    }
  }
}

  3. Map the website’s local agents without guessing syntax

    The website lists these same-PC clients:

    • Claude Code
    • Cursor
    • Codex
    • GitHub Copilot
    • Google Antigravity
    • Hermes Agent
    • OpenClaw
    • OpenCode
    • LM Studio
    • n8n
    • Cline
    • Warp
    • Gemini CLI
    • Grok CLI
    • Windsurf

    For each, the verified connection values are:

    URL: http://127.0.0.1:47831/mcp
Authorization: Bearer <your local API key>

    Use the shared JSON above only where the client accepts that format. The repository does not provide client-specific commands, filenames or schemas for this list. Check your client’s official instructions for its HTTP MCP and header fields; these integrations are setup guidance, not end-to-end certification.

  4. Exclude cloud-only connections

    The website also lists Claude, Lovable, Manus and ChatGPT as cloud agents. There is no functioning loopback configuration for them. Use them for scripts and prompts, then transfer that content manually or through a same-PC agent. Do not publish your app port or treat the website as a cloud authorisation service.

  5. Discover the real tools and test safely

    The server exposes exactly these 14 tool names:

    • flow_status
    • flow_models
    • flow_upload_reference
    • flow_generate
    • flow_get_job
    • flow_wait_job
    • flow_list_jobs
    • flow_cancel_job
    • flow_extend_video
    • flow_edit_video
    • flow_upscale
    • flow_list_characters
    • flow_create_character
    • flow_list_voices

    Start with flow_status and flow_models. Paid generation and follow-up tools require request_id; preserve it on retry. flow_wait_job waits at most 50 seconds per call. Read scope covers inspection; uploads/generation/cancellation need generate; character creation needs characters.

  6. Give the agent the app’s own guide

    Choose “Tải bộ hướng dẫn agent” in Settings. MCP also exposes the guide resource flowstudio://guide. Use its model rules, pricing and retry guidance before allowing generation; check results in Queue and Library.

When it worked

A compatible local client discovers the 14 tools and returns flow_status without starting a generation.

If something goes wrong

Does this prove every listed client works?

No. The repository verifies the server and shared template, not each client’s current config format or end-to-end operation.

Why does the client show no tools?

Check same-PC execution, the current /mcp URL, bearer key and HTTP MCP support. Browser-origin requests are intentionally blocked.