Workflow runtime
Put agent behavior in a workflow
Define model calls, tools, routes, retries, and replay-safe state in versioned TypeScript workflows.
The TypeScript application framework for agents
Write the agent logic.The runtime is already built.
Kortyx packages the TypeScript runtime and React code that kept showing up in every agent project: workflows, streaming, human interrupts, persistent state, resume, rollback, fork, and tracing.
Apache-2.0 frameworkRuns in your appComposable runtime
Runtime connected
// draft-decision.node.ts
const result = await useReason<RefundDecision>({
id: "draft-decision",
model: google("gemini-2.5-flash"),
input: orderAndPolicy,
outputSchema: refundDecisionSchema,
stream: true, emit: true,
structured: {
dataType: "refund.decision",
fields: {
summary: "text-delta",
recommendation: "set",
},
},
});
Schema-validated state · streamed to React
Refund request #4831
Maya Lopez · wireless keyboard
Order
Policy
Draft
Approve
Agent checked the order and refund policy
Customer received the wrong keyboard layout and requested a refund.
Policy check
Within 30 days · unopened
Recommendation
Refund $84.00 to original card
Issue the $84.00 refund?
Execution is checkpointed. The same run resumes after the user decides.
structured streamcheckpointed inputReact state
Defined workflowsStructured streamingHuman in the loopPersistent runtime stateReact clientOpenTelemetry
Why Kortyx exists
I kept rebuilding the same code around every agent.
Every project needed workflow conventions, model hooks, stream parsing, React state, human input, persistence, replay, and tracing.
Kortyx puts that repeated code into a TypeScript framework. You write the workflow, business rules, and interface for your product.
Frontend bridge
Stream text and typed UI state
One SSE protocol carries text, structured data, tool events, interrupts, errors, and abort state into React.
Session semantics
Resume, roll back, and branch correctly
Session checkpoints keep the React transcript and server workflow state together during regenerate, undo, rollback, and fork.
Shared run state
The workflow and React UI read the same run.
Server events map directly to the active React interface. Kortyx runs inside your application and leaves the workflow engine accessible.
04
Your product
Business rules · domain services · product data · interface
Business rules · domain services · product data · interface
03
@kortyx/react
Chat lifecycle · live pieces · interrupts · session actions
Chat lifecycle · live pieces · interrupts · session actions
02
Kortyx server runtime
Workflows · hooks · providers · persistence · transports · telemetry
Workflows · hooks · providers · persistence · transports · telemetry
01
Your infrastructure
Database · model providers · application server · telemetry
Database · model providers · application server · telemetry
A workflow people recognize
A refund request becomes a reviewable decision.
The workflow finds the order, checks policy, drafts a typed recommendation, pauses for approval, then resumes to issue the refund. Each step maps to code and visible interface state.
Walk through the exampleRefund request #4831
MAYA LOPEZ · WIRELESS KEYBOARD
Request
Wrong keyboard layout. Customer asked for a refund to the original card.
Policy match
Delivered 12 days ago · unopened · refund eligible.
Recommendation
Refund $84.00 and close the ticket after confirmation.
Human input
Approve the refund or escalate the request to a specialist.
Included in the framework
The code I needed on every agent project.
useInterrupt()
Human interrupts
Pause for a typed choice, text response, or custom UI. Resume the same run from its checkpoint.
text · set · append
Structured streaming
Stream validated fields into a preview, card, table, or editor while the model is still working.
memory → redis
Runtime persistence
Keep runs in memory locally. Store paused runs and checkpoints in Redis when you deploy multiple workers.
provider neutral
Providers and MCP
Call OpenAI, Anthropic, Google, DeepSeek, Groq, or Mistral. Give each node request-scoped MCP tools and approval rules.
your backend
OpenTelemetry-native
Send model calls, tools, workflow transitions, timing, tokens, and cost to OpenTelemetry or Kortyx Studio.
server → interface
React and Next.js
Connect React through an API route, a buffered Server Action, or a separate Node backend.
Start with two packages
Install the runtime. Add the React client when you need it.
Start on the server with one provider. The quickstart then wires a live Next.js route and the client state needed to render text, structured data, interrupts, and errors.
Follow the Next.js quickstartterminal
# framework + one provider
$ pnpm add kortyx @kortyx/google
# React chat and stream state
$ pnpm add @kortyx/react
01 define a workflow02 expose an SSE route03 render live pieces
Licensing and deployment
Your agent runs in your application.
The framework does not require Studio or a managed control plane. Add visibility when you want it, and keep the execution path under your control.
Provider credentials stay server-side
Prompt, input, and output telemetry is off by default
OpenTelemetry works without Kortyx Studio
Available now
Framework
The TypeScript runtime, providers, React client, CLI, and telemetry API.
Apache-2.0 · you host
Self-hosted preview
Studio
Optional observability for runs, sessions, workflows, interrupts, timing, tokens, and cost.
ELv2 · you host
In development
Cloud
Cloud is in development for teams that want Kortyx to operate Studio.
Details to follow
Common questions
What teams ask before they start.
The short version on integration, deployment, human input, and what is available today.
What does Kortyx replace?
Kortyx replaces the repeated runtime and React integration code around agent workflows: model hooks, streamed events, checkpoints, human input, session actions, and client state. Your business rules and product interface remain yours.
Does Kortyx run inside my application?
Yes. The framework runs in your Node server. Provider credentials stay server-side, and your workflows, services, and product data remain in the application you operate.
Do I need Kortyx Studio or Cloud?
No. The framework works without either. Self-hosted Studio is an optional preview for inspecting runs and sessions. Managed Cloud is in development.
Can I add Kortyx to an existing Next.js or Node application?
Yes. Next.js can connect through a live SSE route or a buffered Server Action. React can also connect to a separate Node backend.
How do human approvals work?
A workflow emits a typed interrupt and checkpoints the active run. React renders a choice, form, or custom component. The response resumes that same run from its checkpoint.
What is open source?
The Kortyx framework is Apache-2.0. Self-hosted Studio uses ELv2. The managed Cloud service is still in development, with commercial details to follow.
Write the workflow for your product.
Kortyx carries stream events, checkpoints, and human responses between the workflow and React.