Use this MCP server to connect to Google Flow, inspect live capabilities, generate and track videos/images, upscale videos, and download exact assets.
Get help and usage examples for Google Flow MCP.
List connected accounts, readiness, and default account.
Diagnose the local Flow Login Bridge.
Begin and complete account connection via Chromium extension and Google account chooser.
Inspect the real Flow workspace for live models, ratios, durations, upscale options, and diagnostics.
Generate videos or images with prompts, references, settings, optional download, and explicit credit authorization.
Poll persistent jobs by ID without resubmitting.
Upscale existing video jobs using available options and credit authorization.
Download a job's exact recorded assets locally with validation and manifests.
Avoid generic browser automation; use the dedicated Flow tools.
Integrates with Google Flow's web generation workspace to read available models and parameters, submit image and video generation tasks, track job progress, download generated assets, and optionally upscale videos using the user's Google Flow account and quota.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Flow Workspace MCP 读取我 Flow 账号可用的视频模型，用横版生成一个纸艺动画镜头，完成后下载到 D:\clips"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
方向 | 内容 |
工作区连接 | 隔离账号会话与实际能力读取 |
任务跟踪 | 同一编号查询，保存分镜组进度 |
素材交接 | 精确匹配原视频并记录下载信息 |
将 Google Flow 接入本地视频制作：持久化任务、复用登录、原视频接口下载。
适合需要持续制作教学视频、科普视频和多个分镜素材的创作者。通过 MCP，AI 助手可以读取账号实际可用的模型与参数、提交生成任务、查询进度，并把对应素材下载到本地。无需另购视频生成 API，但仍使用你的 Google Flow 权限和额度。
开发状态：提供持久化整组分镜调度。后端余额与历史任务额度错误码读取已实测，原视频 HTTP 下载已实测。生成提交仍依赖网页；最新新生成测试在网页设置阶段失败，未提交，纯后端生成尚未完成。此前两段分镜回归在调试恢复后完成，不代表本轮最新生成验收通过。具体证据见 验证记录。多输出和放大未验收。
Related MCP server: gflow-cli
做一条视频往往需要多个片段。重复复制提示词、选择模型和比例、等待生成、确认素材、下载改名，会把制作过程拆成很多人工操作。让 AI 助手逐次通过截图点击，也需要反复读取界面，耗时且容易受页面变化影响。
本项目将这些动作收束为有明确输入、任务编号和输出文件的工具调用。提示词来自你自己的脚本或分镜；Flow 负责生成；本项目负责连接工作区、执行参数、跟踪任务和交付素材。
能力 | 对制作流程的意义 |
复用本地登录会话 | 会话有效时继续工作，减少重复登录 |
原视频优先通过 HTTP 下载 | 已识别的单输出原视频无需进入编辑器、识别下载按钮；接口不可用时回到精确素材匹配路径 |
固定桌面视口 | 避免默认小窗口隐藏 Flow Agent 控件；新建与重连均保持桌面布局 |
记住上次工作区 | 重启后回到已有项目，减少反复创建空项目 |
读取实际模型和参数 | 以账号当前界面为准，不把写死的模型列表当成可用能力 |
持久化任务编号 | 等待、查询和下载围绕同一任务进行，避免排队时重复提交 |
匹配具体生成素材 | 不简单下载图库里“最后一个”或“最新的”文件 |
本地文件与参数记录 | 素材能交给 FFmpeg 或其他剪辑流程，便于按镜头管理 |
截图与控件诊断 | 页面改版或识别失败时留下证据，方便定位问题 |
相比 AI 助手逐步操作浏览器，明确参数和简短任务状态通常能减少界面读取及交互次数；具体 Token 和时间收益需要按同一任务实测。项目不提升视频模型本身的画质，也不减少 Flow 对生成任务收取的额度。
项目内优先调用 GetCredits RPC 读取余额，不需要点击账号菜单；读不到或结构变化仍回退到界面。额度失败优先读取会话中结构化工具结果的 error_code，不依赖中文/英文错误措辞。提交前保存会话轮次基线，只接受新增轮次中唯一精确匹配的请求，避免历史失败误伤当前任务。
目前识别用户额度、模型每日额度（含可升级类别）、工作区账号额度；用户额度码已实测，其余来自 Google 当前客户端分类与自动化测试。明确请求限流不作为额度切号。旧任务无提交前基线时不借用历史错误自动重做；后端读取不支持时保留原有界面路径。私有协议不保证长期稳定，也没有把生成提交迁移成纯 HTTP。详见 接口依据。
flow_download_job 对已追踪素材 ID 的单输出、未放大视频，优先通过 HTTP 获取登录校验参数、查询精确素材并下载原视频，不导航编辑器或点击控件。校验响应素材 ID、下载域名、视频类型及本地媒体信息；记录 downloadTransport=http。无法使用该路径时保留已有网页下载流程，绝不猜测最新素材。
目前是混合实现，不是纯后端生成：生成提交和可用模型配置仍依赖网页；额度与余额优先 HTTP 读取，排队及其他失败仍使用界面路径；登录会话仍由本地浏览器管理。HTTP 读取依赖未公开的网页 RPC，可能随网站更新而变化。独立 HTTP 查询与原片哈希一致性已实测，见 接口验证。
flow_account_credits 优先读取 GetCredits RPC，必要时回退到账号菜单，不生成素材、不购买额度。无法识别时返回 null，不会当成零；当前不推算“今日免费余额”，因为读取的是可用余额，不推断免费和订阅额度的拆分。
先分别连接已有账号，再调用 flow_configure_account_switching，例如 {"accountIds":["personal","backup"]}。顺序为默认生成顺序；生成时显式指定 accountId 则该账号优先。传空数组关闭自动切号。
每个分镜使用 outputs=1。提交前余额明确为零，或当前配置的请求出现明确的额度不足提示时，自动尝试下一个账号，保留同一提示词和参数，并记录实际使用账号与跳过账号。每个账号最多尝试一次。
排队、超时或状态不明的任务不会换号重提。 当前请求对应回复明确失败且提示额度不足（包括 Flow 的“you've reached your credit or daily limit”）时，自动用下一账号重做独立片段，不要求“不扣额度”文字或提前报价。原任务记录替代任务编号；后续查原编号跟踪替代任务。验证码、明确限流、政策拦截、登录失效和原因不明的失败不会触发切号。每个账号最多尝试一次；全部耗尽则停止，不循环切换。
当前 Agent 提交控件未提供可确认的本次生成费用，因此尚未实现按“余额低于报价”跳过账号；不使用旧任务扣费、宣传额度或固定模型价格猜测当前费用。
按分镜逐段提交，每段 outputs=1。例如第 1、2 段已成功，第 3 段明确因额度不足失败：保留前两段，只用备用账号从头重做第 3 段，复制该段提示词、模型、比例、时长、本地参考素材和文件名前缀。原任务与替代任务双向关联，查询原编号可跟踪替代任务，重复查询不会重复提交。不是在另一账号续接已生成的半段视频。
使用 flow_generate_sequence 一次提交有序分镜数组，随后只查询同一个 flow_sequence_status 编号。调度器保留完成段，完成当前段的精确素材下载后自动提交下一段，并沿用成功账号、跳过本组已确认耗尽的账号。results 按分镜编号排列，不按下载时间；每段默认文件名前缀为 shot-001 等。状态查询会继续已授权的生成，并不是只读查询；waitSeconds=0 只读保存的快照。
进度保存在本地 sequences/，重启后查询原编号继续，不重新创建整组。并发查询使用跨进程锁；提交过程中异常退出且尚未保存任务编号时停止并要求检查，不能猜测是否提交成功再重提。调度由状态查询驱动，服务关闭时不会在后台继续生成，不是常驻无人值守队列。不能保证跨账号重生成画面完全一致；Flow 项目不跨账号复制，本地参考文件会复用。
支持 Agent 界面及中文普通生成界面；不同阶段的真实回归结果见验证文档，不能把早期失败记录当成最新状态。
flowchart LR
A[脚本与分镜提示词] --> B[AI 助手调用 MCP]
B --> C[本地 Flow 工作区适配器]
C --> D[Google Flow 生成]
D --> E[任务与素材身份跟踪]
E --> F[本地视频及参数清单]
F --> G[配音、字幕与剪辑流程]连接账号：浏览器扩展将你已有的 Google 登录会话发送到本机回环地址。账号使用隔离的本地浏览器配置；不会要求把密码写入项目。
识别工作区：校验当前 Flow 地址、项目与编辑器控件，读取实际显示的模型、比例和输出数量。适配器同时兼容已有控件和新版 Angular Material 控件。
执行任务：每次生成都有独立任务编号与状态记录。请求必须明确允许消耗额度；项目不购买订阅或额外额度。
确定素材：生成前记录素材基线，之后按素材标识与来源匹配新增输出。匹配不明确时返回错误，避免交付错误视频。
保存结果：下载到指定绝对路径，记录文件大小与 SHA-256。有可用 FFprobe 时，还记录时长、分辨率和编码信息。
生成通过真实网页工作区执行；符合条件的原视频下载优先通过网页自身的 HTTP 接口执行。它不是 Google 官方开发者 API，也不改变 Flow 的额度规则。网页及未公开接口的变化仍可能需要维护。
需要 Node.js 20+、已安装的 Chrome/Chromium，以及可以访问 Flow 的 Google 账号。Windows 是目前实际调试的平台。
git clone https://github.com/QIANLING-0831/flow-workspace-mcp.git
cd flow-workspace-mcp
npm ci
npm run check使用绝对路径注册到 Codex，例如：
codex mcp add flow-workspace -- node "C:\tools\flow-workspace-mcp\dist\index.js"其他支持 stdio 的 MCP 客户端也可以配置 node 加绝对路径 dist/index.js。首次连接步骤见 安装说明。后续先读取账号状态，已有有效连接就直接使用；仅在会话失效、退出登录或更换账号时重新连接。
工具 | 用途 |
| 查询已连接账号与默认账号 |
| 验证工作区、读取可用能力和诊断 |
| 首次连接或重新连接账号 |
| 提交生成请求，返回任务状态 |
| 整组分镜启动一次、按同一编号推进；失败段换号重做后继续下一段 |
| 按原任务编号查询进度 |
| 下载该任务对应的素材 |
| 使用账号实际提供的升级分辨率选项，可能额外消耗额度 |
工具保留 flow_* 名称，以便复用已有客户端和制作流程。
例如对 AI 助手说：
读取我的 Flow 账号可用的视频模型。用我确认的模型生成一个横版纸艺动画镜头，只生成一个输出，不升级分辨率，保存到我指定的素材目录。如果仍在生成，查询同一任务，不要再次提交。
这是素材生成自动化工具，不是完整的脚本、配音、字幕或剪辑软件。
模型、时长、参考素材和升级分辨率是否可用，以账号当前能力为准。
自动化不绕过账号访问限制或验证码，不承诺免费无限生成。
默认保留本地账号配置、任务和诊断；不要将这些运行数据上传到仓库。
生成等待超时不等于任务失败。查询已有任务，避免重复生成和重复消耗额度。
npm run check
npm pack --dry-run本项目由 QIANLING-0831 维护，重点开发接口优先的结果下载、新版工作区适配、会话复用、任务可追踪性和本地视频制作交接。继承代码的来源与版权说明见 SOURCE.md 和 LICENSE。
本项目与 Google 无官方关联。
FIRST CONNECTION STEP. Returns the exact user instructions and extension path immediately; it never blocks waiting for a click. After this response, STOP and tell the user to open Flow Login Bridge in their normal signed-in Chromium, click Connect Flow, and reply when the popup says Session sent. Do not call the completion tool until the user replies. Use this tool for Google Flow; never substitute generic browser/computer-use automation on the Flow website.
Inputs: label, accountId.
SECOND CONNECTION STEP. Call only after the user confirms they clicked Connect Flow and the extension popup says Session sent. Before calling, tell the user that a temporary Google chooser will open and they should click the desired existing account. Requires the begin-step connectionId and explicit user confirmation; if the session was not sent it fails immediately instead of blocking. Use this tool for Google Flow; never substitute generic browser/computer-use automation on the Flow website.
Inputs: connectionId (required), userConfirmedSessionSent (required), label, accountId, +2 more.
The exclusive path to download an existing Flow job. Uses the exact asset identities recorded for that job and fails instead of guessing from gallery order. Never download through generic browser/computer-use tools. Saves to the configured absolute directory, validates files, writes manifests, and does not start a generation.
Inputs: jobId (required).
THE REQUIRED AND EXCLUSIVE PATH for every Google Flow image request. Submits one persistent job, waits briefly, and returns. If status=processing, poll flow_job_status with the SAME job ID; never generate again. It generates or edits through the verified workspace and downloads locally. Never open or control Flow with browser/computer-use tools. Requires explicit credit authorization.
Inputs: prompt (required), outputDirectory (required), confirmCreditSpend (required), model, +8 more.
THE REQUIRED AND EXCLUSIVE PATH for every Google Flow video request. Submits one persistent job, waits briefly, and returns. If status=processing, poll flow_job_status with the SAME job ID; never generate again. It optionally upscales and downloads locally. Never open or control Flow with browser/computer-use tools. Requires explicit credit authorization.
Inputs: prompt (required), outputDirectory (required), confirmCreditSpend (required), model, +10 more.
Call this when the user asks what Google Flow MCP can do, how to use it, or for example requests. Returns a concise multilingual-ready capability guide without opening a browser or spending credits.
Inputs: no named parameters.
MANDATORY before generation. Verifies the real Flow workspace and returns a language-independent live capability map. A landing page returns workspaceAvailable=false and a stop instruction; never browse or scroll it. Use this tool for Google Flow; never substitute generic browser/computer-use automation on the Flow website.
Inputs: accountId.
The exclusive status path for jobs returned by a Flow generation tool. Poll the SAME job ID until completed or failed; never resubmit generation. It safely finalizes any download already authorized by the original request. Never inspect Flow with generic browser/computer-use tools.
Inputs: jobId (required), waitSeconds.
MANDATORY FIRST STEP for Google Flow work. Returns readyForGeneration and the verified default account. When readyForGeneration=true, use defaultAccountId and DO NOT call either account-connection tool. Use this tool for Google Flow; never substitute generic browser/computer-use automation on the Flow website.
Inputs: no named parameters.
Login diagnostic only. Reports whether the localhost Flow Login Bridge is waiting for the Chromium extension. Do not use this for generation and do not open Flow with browser tools. Normally call flow_list_accounts first.
Inputs: no named parameters.
The exclusive path for real Flow video upscaling. Uses live available options and rejects unavailable/upgrade-only choices. Never right-click or control Flow with generic browser/computer-use tools. May consume credits and requires explicit authorization.
Inputs: jobId (required), factor (required), confirmCreditSpend (required), timeoutSeconds.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 11 tools
Generation, job-status, download, and upscale tools target clearly distinct actions on distinct resources. However, three account/workspace-state tools (flow_list_accounts, flow_login_bridge_status, flow_inspect_account) overlap in inspecting readiness, and the two-step connection pair could be confused with them, though the descriptions give strong ordering cues to separate them.
Every tool uses a consistent flow_ prefix with snake_case verb/noun structure (flow_generate_image, flow_list_accounts, flow_download_job, etc.). The pattern is predictable and uniform across all 11 tools.
11 tools is a well-scoped count for a media-generation server. Each tool covers a distinct part of the connect-generate-poll-download lifecycle, and nothing appears redundant or padded.
The core lifecycle is covered: account connection (begin/complete), readiness inspection, image/video generation, upscaling, polling, and downloading. Minor gaps remain, such as no job-history/list operation or job cancellation, which agents may have to work around by retaining job IDs themselves.
Google Flow's agent: chat that plans and makes images and videos, approvals, your own media
LLM chat, text tools, image generation, editing, batch image jobs, and asynchronous video generation
Generate images, video, audio and 3D with FlowNode; results land in your asset library.
Design, save, and run outcome-aligned AI workflows and verifiers, with reliable image output.
