Conversational application platform · PicoFlow
Replace forms, menus, and support flows with AI conversations that understand intent, collect information, execute business logic, and complete real tasks — backed by a runtime built for production.
The flow is the application. The conversation is the UI. PicoFlow is the runtime.
ModelsOpenAI · Anthropic · Google · Azure OpenAI · DeepSeek · Moonshot · Z.AI · Ollama · OpenRouter
SessionsMemory · SQLite · MongoDB · Cosmos DB
RuntimeNode.js 22.5+ · TypeScript · Zod 4 · fits NestJS
The problem with screens
Navigate the menu. Search. Filter. Fill the form. Fix the validation error. Start over. Open a support ticket. Every screen you ship transfers effort to the customer — and every workflow that fails ends in your support queue. A conversational application inverts that: the customer states the intent, and the application does the navigating.
The customer learns your software, or gives up.
The application learns the customer's intent instead.
The runtime underneath
Production conversation needs durable state, validation, business rules, real system calls, concurrency safety, and testing. PicoFlow ships that machinery so your team writes the business process, not the plumbing.
01
One versioned JSON document per conversation holds the active stage, business state, memory, token totals, and structured logs. A customer can leave and come back; the application resumes exactly where the task paused.
02
Define a tool once with a Zod schema, decorate a handler with @Tool, and the runtime dispatches the model's call into your code — where it can hit pricing, booking, CRM, and policy systems with validated arguments.
03
Transitions are return values in typed handlers, not instructions buried in prompt prose. Confirmation before commitment, validation before persistence, and escalation paths are enforced by code, not requested of a model.
04
Memory, SQLite, MongoDB, and Cosmos DB adapters ship with the runtime and implement the same compare-and-swap contract. Swap stores with an environment variable instead of writing another adapter by hand.
05
Turns for one session are serialised in-process, and every store rejects a stale revision with a conflict error. Two browser tabs — or a duplicate request — cannot silently overwrite a customer's in-flight task.
06
Scripted deterministic models assert transitions, persistence, and business state; live scenarios and model judges cover response quality. A conversational application gets regression coverage, not vibes.
Bounded autonomy
Within each stage, the model reasons, asks, and adapts freely. Around it, the flow defines the business envelope: what must be collected, which tools are allowed, what needs explicit confirmation, and what must never happen. A prompt can request “always confirm before booking” — a typed handler enforces it.
stay() keeps the current stage active and hands corrective feedback back to the model.go(Step) moves the durable cursor to the next business stage and seeds its state.direct() answers the customer immediately, with no second model call.
export class ExploreStep extends Step {
getPrompt() {
return `${HotelPrompt.Role}
Collect dates, then call capture_choices.`;
}
defineTool(): ToolType[] {
return [{
name: "capture_choices",
description: "Capture hotel search criteria",
schema: z.object({ json: z.string() }),
}];
}
@Tool
protected async capture_choices(
args: { json: string },
): Promise<ToolResponseType> {
const hotels = await Pricing.search(args.json);
if (!hotels.length) {
return stay("No match. Ask for wider criteria.");
}
return go(PresentStep).withState({ hotels });
}
}
The development loop
PicoFlow changes both sides of the lifecycle. Customers get a conversation instead of screens — and developers describe the application they want, while an LLM generates and iterates the flow and PicoFlow runs and tests it. AI-generated, human-validated application development: the model does the tedious flow-authoring, your team owns requirements, review, and production approval.
Complete applications, not chatbot demos
Each of these is a business process with an outcome — and a flow behind it that collects, validates, calls real systems, and completes. Three ship today as full tutorial applications.
Reservations & commerce
Understand the request, check availability, capture preferences, present priced options, obtain explicit confirmation, and book. Built as a complete tutorial application.
Explore the reservation flowSupport resolution
Verify the customer, inspect account data, execute approved actions with holds, and hand off to a human with a complete case record when needed.
Explore the support flowClaims & document intake
Collect incident details, receive documents, extract structured facts with AI, validate them deterministically, and route exceptions to human review.
Explore the document flowCustomer onboarding
Guide setup in one conversation: gather missing information, verify it, call backend services, handle exceptions, and activate the account.
See the onboarding patternProcurement
Capture purchase intent, apply vendor and policy rules in code, route approvals, create the request, and track status — without a portal.
See the procurement patternYour workflow
If a task takes your customers multiple screens, forms, and a support ticket today, it is a candidate for a conversational application.
Book a walkthroughMore than a chatbot, above an agent framework
| A chatbot | A PicoFlow application |
|---|---|
| Answers questions | Executes business processes end to end |
| Often unstructured | Guided by explicit flows and business rules in code |
| Frequently disconnected from systems | Connected to your APIs, tools, and data with validated arguments |
| Difficult to reproduce and test | Durable sessions, structured logs, and regression tests |
| Best for FAQs and simple support | Designed for real customer tasks and commitments |
And agent frameworks? They orchestrate model and tool calls — useful building blocks, one layer down. PicoFlow claims the application layer: modelling the customer-facing business flow, managing live conversation state, enforcing deterministic business logic, and turning an intent into a completed outcome. For teams comparing directly, we publish an honest, scoped comparison — including the cases where a direct graph framework is the better choice.
@picoflow/core.Adds a CommonJS package entry, exposes the model catalog and last-response utilities, and refines provider and tool-runtime integration.
Source-available, commercially licensed. PicoFlow is not open source. Internal, non-production evaluation is free, including enterprise prototyping; production use requires a commercial agreement. Read the license.
Straight answers
What PicoFlow is, what it needs, and what it does not do yet — including streaming, interrupts, and the licence, because those are cheaper to learn here than after a prototype.
Start here
NodeNext resolution, legacy decorators (experimentalDecorators and emitDecoratorMetadata), and Zod 4. The published 1.1.2 package supports both ESM import and CommonJS require("@picoflow/core") applications. It is designed to sit inside application frameworks such as NestJS. See installation.Evaluating PicoFlow
@picoflow/core depends on @langchain/core and the LangChain provider packages, and a few LangChain types surface in the class you subclass — MessageContent, ToolCall, and DynamicStructuredTool all appear on Step. Persisted memory holds serialised LangChain messages. PicoFlow does not replace LangChain; it adds the durable session, step cursor, tool dispatch, and storage layer above it.interrupt() with a checkpointer is the stronger mechanism. Read the full comparison.Running it in production
Follow the quickstart, then walk through a complete hotel-reservation or customer-support application at your own pace.