Let your assistant drive it¶
gflow-cli is built to be driven by an AI assistant. It ships an MCP server, so Claude, Cursor, or Copilot connects to it natively — no pasting docs, no copying commands back and forth. CLI-to-MCP parity is enforced in CI: anything the terminal can do, your assistant can do.
Start the server¶
gflow mcp run
# serving over stdio:
# • gflow_generate_image
# • gflow_generate_video
# ready — your assistant can drive Flow.
gflow mcp run speaks MCP over stdio; gflow serve exposes an HTTP/SSE endpoint if you'd rather connect over the network. gflow mcp setup helps wire it into a client.
Point your agent at it¶
The repo ships the files an agent reads to understand the tool:
| File | Purpose | Use with |
|---|---|---|
AGENTS.md |
Universal agent onboarding | Cursor, Codex, Aider, Claude Code |
llms.txt |
LLM-readable command reference | Paste into ChatGPT / Claude / Gemini |
CLAUDE.md |
Claude Code memory hub | Claude Code specifically |
Once connected, describe what you want in plain language:
"Make a 9:16 clip of a lighthouse in a storm, moody, slow push-in."
The agent picks the Veo model, aspect, and prompt tooling, runs the generation on your own Flow session, and hands you the files.
What the agent leans on¶
--tool creative-director— expands a terse prompt with Google's 5-component formula before spending credits.gflow instructions— persistent, credit-free brief cards that steer look and tone across generations.- Asset reuse by UUID — chain a still into a video by its id, no re-upload.
--ui-mode— aborts up front when Flow's UI cohort isn't what the command needs, instead of burning a generation.
It runs on your account
Even when an agent drives it, gflow-cli still uses your own headed Flow session and bills your own credits. The same alpha/account-risk caveats apply.
Next: Known issues →.