Detailed Analysis
A Reddit developer's discovery of unused Veo 3.1 video credits has produced a practical case study in how MCP (Model Context Protocol) servers are quietly expanding what Claude can automate on a user's behalf, even against services that offer no public API. The core insight is mundane but telling: Google's AI Pro subscription bundles 1,000 monthly credits for Flow, its Veo 3.1 video generation tool, yet the interface is a slow, conversational approval flow poorly suited to producing more than a clip or two before becoming tedious. Rather than accept that friction — or keep paying for a separate video-generation subscription — the developer built google-flow-mcp, an open-source (MIT-licensed) MCP server that lets Claude drive Flow directly: batching clip requests, tracking spend against a budget ceiling, and submitting renders concurrently instead of one at a time.
The technical details matter because they illustrate a recurring pattern in the current MCP ecosystem: agents are increasingly being pointed at consumer web apps that were never designed to be automated, using a combination of reverse-engineered API calls and UI automation as a fallback. Here, the author discovered that Flow's frontend is built on tRPC, meaning there's a real (if undocumented) API underneath the chat-style interface, which the MCP server calls directly when possible and only falls back to simulated clicking when no endpoint exists. This is a more durable architecture than pure browser-automation scrapers, which tend to break the moment a site changes its CSS. The project also surfaces a genuinely useful UX bug report: Flow's "Approve" button sits next to an "Approve, do not ask again" option with nearly identical inner text, a trap for any naive automation script that could result in silent, budget-busting charges. Building price verification before every approval — and making rejection free — reflects the kind of safety-conscious design that agentic tooling increasingly needs when it's handling real money on a user's behalf.
Beyond the specific hack, the piece reflects a broader shift in how power users relate to Claude: not as a chatbot to converse with, but as an orchestration layer sitting on top of other products' interfaces, closing gaps that vendors themselves haven't bothered to fix. Google Flow has no batch mode and no official API for third-party tools, so instead of waiting for Google to ship one, an individual developer used Claude via MCP to build a bridge — turning an underused subscription perk into an actual production pipeline, paired with free open-source tools like ffmpeg, DaVinci Resolve, and faster-whisper for the assembly work Flow itself doesn't do. This "glue layer" use case is becoming one of the most common real-world applications of Claude's tool-use capabilities: not generating content itself, but coordinating and optimizing workflows across services that already exist.
The caveats the author includes are equally instructive about where this class of tool currently stands. It's explicitly v0.1, uncalibrated against a live Flow session, and dependent on UI selectors that may already be stale. It also operates in a legal and technical gray zone — automating a Google product through an unpublished interface, riding on a user's already-authenticated Chrome session, with no guarantee Google won't change something and break it overnight. That fragility is characteristic of the current wave of MCP-based automation tools: genuinely useful, often built quickly by a single motivated user, but resting on unofficial foundations that could vanish the moment a platform vendor decides to lock things down. As Anthropic and others continue investing in the MCP standard as a way to let Claude act across third-party software, this kind of community-built connector — solving a real, if small, financial and workflow problem — is a preview of how much of that ecosystem is likely to be built: bottom-up, informally, and slightly ahead of what platform owners have sanctioned.
Read original article →