AI Worksflows with LiveBindings and Replicate

Build visual AI workflows in Delphi FMX — wire Replicate.com models together in the LiveBindings Designer, press Run, and watch the chain execute.

---

Drop model nodes on a form, drag a line from one model's output to the next model's input, and the workflow runs itself: generate an image with flux-schnell, and the upscaler downstream launches automatically the moment the URL lands in its input. Inputs are typed from each model's JSON schema, outputs fan out to grids and images, and a JSON-aware relay picks which field of an output travels on down the chain.

┌─────────────────┐      ┌──────────────────┐      ┌─────────────────────┐
│  flux-schnell   │─────▶│  > output[0] >   │─────▶│ recraft-crisp-      │
│  "a red car..." │      │  (TReplicateRelay)│      │ upscale   AutoRun   │
└─────────────────┘      └──────────────────┘      └─────────────────────┘
   TReplicateNode          JSON field picker           TReplicateNode
        └──────────── all three are LiveBindings Designer lines ───────────┘

Contents

• Why this project
• Components
• Quick start
• Your first workflow
• Samples
• Repository layout
• Requirements
• Notes & caveats
• License

Why this project

• Visual, not glue code. Model→model chains are drawn in the LiveBindings Designer, so the topology of a workflow is visible on the form — no hidden event wiring.
• Schema-driven. Inputs are typed straight from each model's JSON schema; TReplicateInputPanel builds a full input form (switches, sliders, enum-aware combos, file pickers) at runtime with zero configuration.
• Cascading execution. Set AutoRun and a downstream stage launches itself the instant its input arrives — one node fanning into the next into the next.
• Code-first when you want it. Every visual capability has a plain-code path (Chat.PipeTextTo(Gen).PipeImageTo(Up)) for pipelines without bindings.
• Mixed providers. An optional SmartCore chat engine (OpenAI / Claude / Gemini / Ollama) plugs into the same nodes and chains as the Replicate models.

Components

Component	Kind	Purpose
TReplicateNode	control	The workflow node. Visual box (model name + live status) with bindable InputPrompt/InputImage/Output*/OutputJSON members. Designer-drawable node→node chains; AutoRun cascades stages.
TReplicateRelay	control	The valve. JSON in, InputPath (output[0], output.text, …) picks the field, extracted value out. Couples anything to anything, including REST-component bind sources.
TReplicateSchemaLink	component	Schema→schema workflow edge: Mappings of input_name=output_path, applied atomically on completion, then launches the target. Design-time Map Schemas... dialog.
TReplicateInputPanel	control	Builds an input form from the model's JSON schema at runtime: switches, sliders, combos (enum-aware), file-browse edits — zero configuration.
TReplicateOutputItemsSource	bind source	Multi-output models normalized to rows (ItemIndex/Name/Value) — bind grids and list controls to a model's full output array.
TReplicateDataBindSource	bind source	Dataset-backed single-model component: per-instance schema fields, designer field links, grid-friendly.
TReplicateModel	component	The Replicate engine (REST, polling, schema mapping, notifications). Powers the node; usable standalone for code-wired pipelines.
TAIEngineNode	control	A node face for any engine dropped on the form (Engine reference, the TDataSource.DataSet idiom) — the extension point for non-Replicate providers.
TSmartCoreChatEngine	component	Optional (RAD Studio 13+): chat engine over Embarcadero's SmartCore AI pack (OpenAI/Claude/Gemini/Ollama drivers) — mixed-provider workflows via the same nodes and chains. Ships in the separate LiveAISmartCore package.
TReplicateBindSource	bind source	The original single-model bind source (kept for compatibility).

Design-time extras: a model picker (searches Replicate's catalog), Load Model Schema (caches the schema into the form file), Map Schemas..., and an experimental Generate Input Form verb that emits controls + field links onto the form.

Quick start

1. Build & install the packages (order matters)

• Open Packages/LiveReplicate.dproj → Build (runtime).
• Open Packages/dclLiveReplicate.dproj → Build, then Install (design-time).
• Restart the IDE. A Replicate palette page appears.

2. Set your token

Preferred: the REPLICATE_API_TOKEN environment variable — the components and all samples fall back to it.

Warning

Never commit a token. Form files are code, and a pasted token is saved with them. Get a token at replicate.com/account/api-tokens.

3. Run a sample

Start with Samples/08-Nodes for the node workflow, or see the samples index for the full learning path.

Your first workflow, from scratch

1. Drop two TReplicateNodes on an FMX form. Right-click each → Pick Replicate Model... (e.g. black-forest-labs/flux-schnell and an upscaler), then Load Model Schema.
2. Open the LiveBindings Designer and drag NodeGen.OutputImage → NodeUpscale.InputImage.
3. Set NodeUpscale.AutoRun := True.
4. In a button click:

   NodeGen.SetInputValue('prompt', Edit1.Text);
   NodeGen.Run;

Stage 2 launches itself when stage 1 completes. Bind Status, Logs, and OutputImage to controls for live progress; use a TReplicateRelay between nodes when you need to pick a specific field out of the output JSON.

Samples

The Samples/ folder is a numbered learning path — each sample is self-contained: open the .dproj, set REPLICATE_API_TOKEN, run.

#	Sample	Component(s)	Demonstrates
01	Console	TReplicateBindSource	The service layer with no UI and no bindings — the diagnostic harness
02	BindSource	TReplicateBindSource	The original single-model FMX demo (legacy, kept for compatibility)
03	DataBindSource	TReplicateDataBindSource	Dataset-backed bind source: per-instance schema fields, designer field links, grids
06	ModelChain	TReplicateModel	Code-wired chaining — no .fmx bindings, no designer; the code-first pipeline pattern
08	Nodes	TReplicateNode	The visual node workflow — designer-drawn model→model lines with AutoRun cascade
09	Schema	TReplicateInputPanel, TReplicateSchemaLink, TReplicateOutputItemsSource	Schema-built input form, schema mapping edge, output items grid
10	Relay	TReplicateRelay	The JSON-field-picking valve: OutputJSON → relay(output[0]) → InputImage
11	SmartFlow	TSmartCoreChatEngine, TAIEngineNode	Mixed providers: Claude writes the prompt → flux-schnell → upscaler in one cascade. Requires RAD Studio 13+ + SmartCore.
12	PureBindings	TReplicateNode	Zero code. gpt-oss-20b → flux-schnell → upscaler with an empty form unit — the whole pipeline in bindings
13	CodeFirst	TReplicateModel	Zero bindings. The same pipeline wired in code: Chat.PipeTextTo(Gen).PipeImageTo(Up)

Start with 08-Nodes, then 09-Schema and 10-Relay — those three are the current architecture. 12-PureBindings is the payoff: a three-stage pipeline with not a single event handler in the unit. Numbering gaps are deliberate — they belong to superseded approaches, preserved under Samples/archive with notes on why each was outgrown.

Repository layout

Source/     Runtime components (one unit per component)
Packages/   LiveReplicate (runtime) + dclLiveReplicate (design-time)
Samples/    Numbered learning path (superseded approaches under Samples/archive)

Requirements

• Delphi 13 with FireMonkey (also builds on Delphi 12.x Athens)
• FireDAC — in-memory tables only, no database
• A Replicate API token
• Internet access at runtime
• For mixed-provider workflows: the SmartCore AI pack (from GetIt) for the TSmartCoreChatEngine and the SmartFlow sample

Notes & caveats

• Developed against Delphi 13 and Delphi 12.x. Designer-integration behaviors are version-sensitive.
• Predictions run on Replicate's infrastructure and cost money — the components guard against accidental design-time launches, and AutoRun only fires on non-empty inputs, but review chains before running them.
• Win32/Win64 project targets are configured; other FMX platforms should work but are untested.

Origins

The initial project was scaffolded in RAD Studio Kai, Embarcadero's AI-assisted development for Delphi, and has been developed further since.

License

Released under the MIT License — © 2026 FMXExpress.