🎬 google-flow-mcp
|
接入你的智能体
发布者自己的命令。不是 PlugMCP 端点。
用我们的密钥与日志托管一份副本。
npx -y @park-sang-gwon/[email protected]Claude Code 没有针对本地命令的安装链接。复制下方命令并在终端中运行。
claude mcp add park-sang-gwon-google-flow-mcp -- npx -y @park-sang-gwon/google-flow-mcp@0.3.1在终端中运行。
{ "mcpServers": { "park-sang-gwon-google-flow-mcp": { "command": "npx", "args": [ "-y", "@park-sang-gwon/[email protected]" ] } }}claude_desktop_config.json
远程 URL 还是本机命令? 两种安装方式的区别
Let your AI assistant make videos in Google Flow (Veo, Nano Banana) — and stitch them into long, seamless takes with Scene Builder — in plain language:
- "Generate an 8-second vertical clip from this keyframe: a rooster crowing on a stone wall at dawn."
- "Add it to a scene and extend it twice — keep the slow push-in going, then let the camera rise into the sky."
- "Make three 9:16 poster images of a copper weathervane with Nano Banana."
- "Show me what the Flow page looks like right now." (
flow_inspect/flow_screenshot)
This is a Model Context Protocol (MCP) server that drives Google Flow through your own Chrome session. It works with Claude Code, Claude Desktop, Cursor, VS Code, Gemini CLI — any MCP client — and with plain JSON-RPC scripts.
flowchart LR
C["🤖 MCP client<br/>Claude Code · Claude Desktop · Cursor · VS Code · Gemini CLI<br/>or any JSON-RPC script"]
subgraph S["google-flow-mcp — runs on your machine"]
direction TB
T["11 tools<br/>session · generate · media · scene builder"]
F["Flow automation<br/>agent composer · scene timeline · media ids"]
J["Job records<br/>~/.google-flow-mcp/jobs"]
end
subgraph B["Google Chrome (your profile, port 9333)"]
P["one tab owned by this server"]
end
G["☁️ Google Flow<br/>flow.google.com · Veo 3.1 · Omni · Nano Banana"]
C <-->|"MCP over stdio"| T
T --> F
F <-->|"Playwright over CDP"| P
P <-->|"HTTPS, your login"| G
F -.-> J
Chrome is launched outside Playwright (so Google's sign-in accepts it) with a dedicated profile that keeps your login. The server never sees a password; prompts go to Flow exactly as you wrote them; downloads come back through the browser's own authenticated session.
resume: true finishes them after a restart or timeoutflow_inspect): snapshot/diff of the live page, text search, media tile positions, network watch{ ok, code, message, details } — clients branch on error codes, never on message textdoctor checks Chrome, the port, your login and finds hung tabs that would block connectionsclaude mcp add google-flow -- npx -y @park-sang-gwon/google-flow-mcp
claude_desktop_config.json → mcpServers:
{
"mcpServers": {
"google-flow": { "command": "npx", "args": ["-y", "@park-sang-gwon/google-flow-mcp"] }
}
}
~/.cursor/mcp.json (or project .cursor/mcp.json):
{
"mcpServers": {
"google-flow": { "command": "npx", "args": ["-y", "@park-sang-gwon/google-flow-mcp"] }
}
}
.vscode/mcp.json:
{
"servers": {
"google-flow": { "type": "stdio", "command": "npx", "args": ["-y", "@park-sang-gwon/google-flow-mcp"] }
}
}
~/.gemini/settings.json:
{
"mcpServers": {
"google-flow": { "command": "npx", "args": ["-y", "@park-sang-gwon/google-flow-mcp"] }
}
}
git clone https://github.com/ParkSangGwon/google-flow-mcp && cd google-flow-mcp
npm install && npm run build
claude mcp add google-flow -- node "$PWD/dist/cli.js"
npx -y @park-sang-gwon/google-flow-mcp doctor
The first run launches a dedicated Chrome (~/.google-flow-chrome). Sign in to Google in that window — once. doctor then reports flow: logged in. From now on the server reuses that profile; there is nothing to configure unless Chrome is somewhere unusual (see Configuration).
Ask your assistant to open a Flow project and prepare a video without spending credits:
"Call
flow_project_openon https://flow.google.com/project/… thenflow_generate_videowithauto_confirm: falsefor an 8-second 9:16 clip of …"
You get status: "ready_for_confirmation" and a screenshot path. Say "go ahead" and the agent repeats the call with auto_confirm: true.
sequenceDiagram
autonumber
participant A as 🤖 Agent
participant S as google-flow-mcp
participant B as Chrome tab
participant F as ☁️ Google Flow
A->>S: flow_generate_video {prompt, model, reference_images, auto_confirm}
S->>B: open project · new agent session · settings (ratio, count, model, "confirm: never")
S->>B: attach keyframe → verify thumbnail (else REFERENCE_NOT_ATTACHED, nothing sent)
S->>B: fill instruction
alt auto_confirm = false
S-->>A: status: ready_for_confirmation + screenshot (0 credits)
else auto_confirm = true
S->>S: write job record (baseline tile count)
S->>B: send (Enter) → handle approval card / policy refusal
B->>F: generate
loop every 6 s, up to 30 min
S->>B: more tiles than the baseline?
end
F-->>B: clip ready
S->>B: authenticated GET flow.google.com/asb/<token>
S-->>A: status: completed, files[], media_ids[], job_id
end
If the server dies or times out during the loop, call the same tool again with resume: true — it reloads the job, skips the composer and only waits/downloads.
Flow's Scene Builder extends a clip by analysing its last frames and generating what happens next, keeping motion and ambient sound. This server automates the whole path:
flowchart LR
K["🖼️ keyframe"] --> V["flow_generate_video<br/>Veo 3.1 Fast · 8 s · 10 credits"]
V --> A["flow_scene_add<br/>tile ⋮ → Add to Scene → Create scene<br/>0 credits"]
A --> E1["flow_scene_extend #0<br/>Veo 3.1 Lite · +7 s · 5 credits"]
E1 --> E2["flow_scene_extend #1<br/>+7 s · 5 credits"]
E2 --> N["…"]
V -.->|"seed.mp4"| X["ffmpeg concat<br/>(local, free)"]
E1 -.->|"hop1.mp4"| X
E2 -.->|"hop2.mp4"| X
X --> O["🎬 22 s take, one continuous motion"]
What we measured on the live UI (September 2026, details in docs/scene-builder.md):
[!TIP] Restate the subject, wardrobe, light and camera move in every hop prompt — Flow does not carry the previous prompt over.
Recipe:
seed = flow_generate_video { model: "veo-3.1-fast", duration: 8, reference_images: ["/abs/keyframe.jpg"], ... auto_confirm: true }
scene = flow_scene_add { project_url, media_id: seed.media_ids[0] }
hop1 = flow_scene_extend { scene_url: scene.scene_url, prompt: "…continues…", after_clip_index: 0, output_dir, auto_confirm: true }
hop2 = flow_scene_extend { ..., after_clip_index: 1, ... }
ffmpeg -i seed.mp4 -i hop1.mp4 -i hop2.mp4 -filter_complex "[0:v][0:a][1:v][1:a][2:v][2:a]concat=n=3:v=1:a=1" take.mp4
Full input/output schemas: docs/tools.md (generated from the code, always in sync).
Every tool answers with one JSON text block:
// success
{ "ok": true, "status": "completed", "files": ["/abs/out/flow_ab12cd34_job.mp4"], "media_ids": ["ab12cd34"], "job_id": "mtk0…" }
// failure (isError: true) — never thrown, always structured
{ "ok": false, "code": "REFERENCE_NOT_ATTACHED", "message": "0/1 reference images attached; nothing was sent", "recoverable": false, "details": {}, "screenshot": "/…/reference-not-attached.png" }
Defaults work on macOS. Override with ~/.config/google-flow-mcp/config.json (npx -y @park-sang-gwon/google-flow-mcp init writes one) or FLOW_MCP_* environment variables:
The server speaks MCP over stdio; a raw JSON-RPC client is ~40 lines:
import json, subprocess, itertools
proc = subprocess.Popen(["npx", "-y", "@park-sang-gwon/google-flow-mcp"], stdin=subprocess.PIPE, stdout=subprocess.PIPE, text=True, bufsize=1)
ids = itertools.count(1)
def rpc(method, params=None, notify=False):
msg = {"jsonrpc": "2.0", "method": method, "params": params or {}}
if not notify: msg["id"] = next(ids)
proc.stdin.write(json.dumps(msg) + "\n"); proc.stdin.flush()
if notify: return None
for line in proc.stdout:
if line.startswith("{") and json.loads(line).get("id") == msg["id"]:
return json.loads(line)["result"]
rpc("initialize", {"protocolVersion": "2024-11-05", "capabilities": {}, "clientInfo": {"name": "me", "version": "0"}})
rpc("notifications/initialized", notify=True)
res = rpc("tools/call", {"name": "flow_status", "arguments": {}})
print(json.loads(res["content"][0]["text"]))
Only JSON-RPC frames are written to stdout; logs go to stderr and ~/.google-flow-mcp/logs.
Logs: ~/.google-flow-mcp/logs/server-YYYY-MM-DD.log · screenshots of every phase and failure: ~/.google-flow-mcp/screenshots/.
Is this an official Google product? No. It is an unofficial automation of the Flow web UI; it uses no private API and does nothing you could not do by hand in the same browser.
Does it store my Google password? Never. You sign in inside Chrome; the session lives in the Chrome profile directory. Treat ~/.google-flow-chrome like a password store.
Can it download the finished scene from Scene Builder? Not in the current Flow build — the export renders but no file reaches an automated tab. Download the clips by media_id and concatenate locally (the seam is frame-exact).
Where is "Jump to"? The current Scene Builder menu only offers "Add clip" and "Extend", so there is nothing to automate yet.
Which Flow UI languages work? Korean and English are matched; other languages need their strings in src/flow/labels.ts — flow_inspect shows you exactly what is on screen.
npm install
npm run check # format · lint · typecheck · unit + stdio contract tests · docs freshness
npm run build
npx tsx scripts/call.ts flow_status # run one tool over real stdio
See docs/development.md for the flow_inspect discovery loop and docs/architecture.md for the design. Live tests that touch Flow are gated behind FLOW_E2E=1. Please follow Conventional Commits (CONTRIBUTING.md).
MIT. Google Flow, Veo and Nano Banana are trademarks of Google LLC; this project is not endorsed by or affiliated with Google.
Unofficial MCP server that drives Google Flow (labs.google Veo / Nano Banana) through your own Chrome session — video & image generation plus Scene Builder (extend, jump to, download)
{ "mcpServers": { "park-sang-gwon-google-flow-mcp": { "command": "npx", "args": [ "-y", "@park-sang-gwon/[email protected]" ] } }}工具
暂无工具列表。远程服务器在首次健康检查后会获得工具列表。
常见问题
在本页复制适用于 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 配置。它使用发布者的安装方式,而不是 PlugMCP 代理。
此记录未列出必需的环境变量。
@park Sang Gwon/google Flow 通过 stdio 在本地运行。
上次检查时间:2026-11-16。 Unofficial MCP server that drives Google Flow (labs.google Veo / Nano Banana) through your own Chrome session — video & image generation plus Scene Builder (extend, jump to, download)