FlutterFlow MCP: the offline stocktake test
FlutterFlow MCP helped build an editable stocktake app, but its first device launch still lacked the AI feature. Stuart reports offline image classification working after further Codex and native-code repairs. The opportunity is faster app development with a visual handover, provided the actual feature is checked on the target device.
This review follows Stuart's 16:54 FlutterFlow MCP episode from May 2026. The recording date is unknown. Findings below are reported from the full captions, not independently reproduced device tests. The introduction announces three experiments; this episode covers the first, not three completed builds.
Where could this help a business?
An inventory task is a useful starting point: photograph an item, suggest what it is, then save a record. A shop stocktake or field-service equipment log would be illustrative adaptations, not measured client results. Agree the labels, review process and acceptable mistakes before replacing a person's judgement.
The earlier discussion also suggests using an agent for fiddly changes and bug fixes in existing apps. That is a proposed use case here, not a demonstrated test of every existing project. Existing-app ideas, 2:18-2:30.
What does MCP connect?
MCP lets a coding agent use FlutterFlow tools. In the recording, Stuart starts the FlutterFlow CLI, supplies an API token and opens the generated workspace in Codex. He reports automatic registration with several coding environments and a scaffold containing FlutterFlow-specific Dart references. The agent then works through a project-editing layer rather than simply changing exported app files.
The useful distinction is the editable FlutterFlow project. Current FlutterFlow AI CLI documentation describes a local workspace connecting an agent to FlutterFlow's cloud project APIs. The project remains the source of truth. Building that project through MCP and running its finished AI feature offline are separate operations.
CLI, setup and workspace, 2:37-5:00; Project-editing workflow, 6:28-6:54.
What was the stocktake brief?
A visual inventory tool: take a photo, identify its contents with an on-device model, and save a local record. It was intended to work offline after model preparation, rather than send every image to a hosted chatbot. Stuart calls it a dedicated visual inventory tool
. Camera access, local storage, a model download and real native inference were essential.
The commentary names Google's Gemma 4 and LiteRT-based inference. Automatic captions garble some model and SDK terms, so this article does not establish the exact model artifact or quantisation. A working camera screen alone would not satisfy the brief. Prompt and inventory task, 5:04-6:22.
What appeared in FlutterFlow?
Stuart reports generated pages, reusable components, a history screen and custom-code blocks appearing in the visual builder. That is more useful for handover than an isolated mock-up. He also explicitly stops short of calling it a general Flutter-to-FlutterFlow converter.
The design was basic and fell short of the promotional examples in his view. He suggests refining the design in FlutterFlow Designer and returning it to FlutterFlow, instead of accepting the first output. The generated project also overrode path_provider 2.1.4 with 2.1.5 and used custom SQLite logic. Those are recorded dependency choices, not versions we recommend for a new project; overrides and generated database code still need review.
Generation, design and dependencies, 6:58-9:01; Design refinement discussion, 10:17-10:54.
Why wasn't a running app enough?
The episode reports an APK build and an Android emulator running the interface despite compiler warnings. Its model-preparation step seemed too quick for the roughly 1.5 GB download Stuart expected. That figure is his stated expectation, not a measured download size or verified current model requirement.
He next describes an iPhone X build, a provisioning-profile adjustment in Xcode, and installation. The camera captured an image and saved a database record, but the crucial classification metadata was absent. Launching the app proved less than delivering the promised feature.
Android build and preparation concern, 9:07-10:16; iPhone attempt and missing classification, 11:11-12:41.
What still needed native code?
The agent's output identified missing host-side method-channel implementations. A method channel carries messages between Dart and platform code; the app needs the corresponding native handler to perform the requested operation. Calling an empty bridge cannot produce the missing AI result. This distinction agrees with Flutter's platform-channel documentation.
Stuart asks Codex to finish the bridge rather than write it himself. The episode describes replacing stubs with native calls and a Python script to inject Android bridge code before a local build. He flags the boundary becoming murky between the coding agent and FlutterFlow MCP. This repair should not be credited to MCP alone, nor treated as a recipe tested across both platforms.
Missing native implementation and repair, 12:49-14:48. FlutterFlow now also documents native SDK integration using method channels; use an existing compatible plugin where it already solves the task.
What result does Stuart report?
After further debugging, Stuart reports a stocktake app that takes a photo, processes it on-device and returns its contents offline. He says classification takes a while and puts the whole experiment, including fixes, at around an hour and a half. His comparison with an earlier months-long experiment is personal experience, not a controlled benchmark or delivery estimate.
The captions do not independently prove flight-mode operation, model identity, classification accuracy, latency, or a separate successful Android and iOS result. We have not reproduced this on a physical device. His prediction of large performance gains in six months is a prediction, not a measured outcome.
What has changed since the recording?
Checked 8 October 2026: current CLI documentation uses flutterflow ai init, supports new or existing projects, and describes project-scoped agent registration. It separates project edits from running, deployment and store submission; Live Sessions require a running app connected through the desktop workflow. These current instructions supersede treating the recorded setup as a universal installation guide.
The recording's macOS-only comment is dated. The current FlutterFlow AI product page advertises its desktop Agent Panel for macOS and Windows; that is a different surface from an arbitrary hosted MCP client. Check your intended workflow's prerequisites. Google's current LiteRT-LM overview names an on-device inference runtime, but does not verify this recorded app's model or bridge.
What should we verify before handover?
- The actual task: a photo creates useful, reviewable inventory data, including incorrect and uncertain classifications.
- The target devices: test Android and iOS separately where required, including model download, offline use, restart and slow inference.
- The whole build: review dependency overrides and retain any native bridge steps so a clean rebuild preserves the feature.
- The editable project: confirm who maintains the FlutterFlow project, generated code and native changes.
This is our practical checklist drawn from the experiment, not a claim that every item was tested in the video. Start with one useful workflow and prove it before expanding the scope.
Sources and evidence limits
Primary source: Stuart Gardoll, Let's Build, FlutterFlow MCP is here. Worth the wait? We SERIOUSLY put it through its paces., 16:54. Retrieved metadata lists an upload on 1 May 2026; the public page displays a premiere on 2 May. The original recording date is unknown. This article was prepared on 8 October 2026 from the full exact-video English auto-captions, with current official references checked separately.
Automatic captions contain transcription errors. Reported demonstrations remain narration, not independently verified device or provider functionality. The final section announces this as part one of three and mentions project/native-code links; we have not imported or executed them. The unrelated owner file MCP.srt covers BuildShip and was not used. Series scope and closing, 16:08-16:42.