Why not build it ourselves?

You can build an agent with any agent framework. What takes months is everything around it: sign-in, roles, credentials, tracing, deployment and operations. Operaide brings all of that, so your team writes only the business logic and the first app runs in days. The apps stay plain TypeScript in your Git, on your own infrastructure.

Which AI models can we use?

Commercial, European and self-hosted AI models: OpenAI, Anthropic, Mistral, Azure, Google, Amazon Bedrock, OpenRouter and more, or open-source models via Ollama or LM Studio. Your administrators configure them, and every step of an app can use its own AI model, chosen for its task.

How much TypeScript do I need?

Enough to read and review code. AI coding agents write much of it, but someone on the team should be able to judge what they produce. Familiarity with Zod helps.

Which AI coding agents can I use?

Claude Code and Codex are set up in Operaide Studio, and Operaide Code is our own terminal agent. Other assistants work too: the Reaktor framework reference is public in our docs.

Can I develop locally instead of in the browser?

Yes. Download an app template, work in your own editor and deploy with the same commands. Operaide Code runs on macOS, Linux and Windows with your own AI provider keys.

Can I use any npm package?

Apps run in a sandbox with a fixed set of modules, including axios, zod, the Vercel AI SDK, pdf-lib and the MCP SDK, plus any published operaide-* library. This keeps apps secure and predictable across tenants.

Agentic AI as code.In TypeScript.

Compose agents from small, typed steps. Let Claude Code, Codex or Operaide Code do the typing. The platform handles sign-in, permissions and operations.

Apply for the Community Edition Read the docs

Operaide Studio/ invoicesIn the browser · no local setupPreview

extract-invoice.reaktor.tssteps.ts

invoicesextract-invoice.reaktor.tsInvoice

1import { z } from 'zod';
2import { registerReaktorDefinition } from '@operaide/aktor';
3import { aktorExtractInvoice } from './steps';
4 
5// The fields the AI model must return
6const Invoice = z.object({
7  number: z.string(),
8  total: z.number(),
9  dueDate: z.string(),
10});
11 
12registerReaktorDefinition({
13  reaktorDefinitionId: 'extract-invoice',
14  outputSchema: Invoice,
15  aktor: aktorExtractInvoice,
16});

ProblemsOutputTerminal

$npm run check

✓Types checked

✓Follows the Reaktor framework

✓Preview ready on port 3000

Operaide Code

 ██████╗ ██████╗ ███████╗██████╗  █████╗ ██╗██████╗ ███████╗
██╔═══██╗██╔══██╗██╔════╝██╔══██╗██╔══██╗██║██╔══██╗██╔════╝
██║   ██║██████╔╝█████╗  ██████╔╝███████║██║██║  ██║█████╗  
██║   ██║██╔═══╝ ██╔══╝  ██╔══██╗██╔══██║██║██║  ██║██╔══╝  
╚██████╔╝██║     ███████╗██║  ██║██║  ██║██║██████╔╝███████╗
 ╚═════╝ ╚═╝     ╚══════╝╚═╝  ╚═╝╚═╝  ╚═╝╚═╝╚═════╝ ╚══════╝
CODE

Ask Operaide Code to change the app…

extract-invoice.reaktor.tsAsk before editsGLM-5.2

main0 problemsLn 7, Col 34TypeScriptOperaide Code

Why developers like it

Less plumbing. More agents.

The platform handles everything around your apps. You write the logic and ship it the same day.

TypeScript your coding agent already knows

Typed steps in plain TypeScript, in your Git. Coding agents write it from the first prompt.

No auth, no tenants, no ops code

Sign-in, roles, tenants, credentials and deployment come with the platform.

Your security team will say yes

Your infrastructure, local or European AI models. No stalling in the data protection review.

What an agent looks like

A few lines of TypeScript. A production API.

This agent extracts data from an invoice. The schema guarantees a typed result, the registration makes it a REST endpoint.

Typed output

total: z.number()

The Zod schema defines the result. No JSON parsing.

AI model per deployment

aktorAISettingProviderModel()

Provider and AI model are a setting, not code.

Instant API

POST /reaktors/extract-invoice

Registered agents become endpoints, protected by API keys and roles.

src/extract-invoice.reaktor.tsTypeScript

1import {2  aktorAIGenerateObject,3  aktorAISettingProviderModel,4  aktorMessagesFromSystemAndUser,5} from '@operaide/ai';6import {7  aktorConst,8  createAktorComposition,9  registerReaktorDefinition,10} from '@operaide/aktor';11import { z } from 'zod';12 13const Invoice = z.object({14  number: z.string(),15  total: z.number(),16  dueDate: z.string(),17});18 19const aktorExtractInvoice = createAktorComposition(20  'aktorExtractInvoice',21  ({ document }) =>22    aktorAIGenerateObject({23      messages: aktorMessagesFromSystemAndUser({24        system: aktorConst('Extract the invoice fields.'),25        user: document,26      }),27      providerModel: aktorAISettingProviderModel(),28      schema: aktorConst(Invoice),29    }),30);31 32registerReaktorDefinition({33  reaktorDefinitionId: 'extract-invoice',34  inputSchema: z.object({ document: z.string() }),35  outputSchema: Invoice,36  aktor: aktorExtractInvoice,37});

Call it from any system

TerminalShell

1curl -X POST https://operaide.example.com/api/v2/orgs/$ORG/apps/invoices/reaktors/extract-invoice \2  -H "X-API-KEY: $OPERAIDE_API_KEY" \3  -H "Content-Type: application/json" \4  -d '{ "document": "…" }'

Workflow

From new app to production.

If you know npm and Git, you know the workflow.

Start

Create a new app, then choose: Operaide Studio or your own tools.

Run

Deploy with one command. Test agents in the browser, step by step.

Publish

Review in Git, release to the app registry. Admins install it in production.

TerminalShell

1npm install2npm run check     # ESLint, Prettier, TypeScript3npm run deploy    # to a development instance4npm publish       # release to the app registry

Reaktor framework

Building blocks for common agent tasks.

Stable, backward-compatible packages cover what most agents need.

@operaide/aktor: The core: compose steps, branch and loop, settings, roles and permissions, scheduled agents, logging.
@operaide/ai: AI model calls across providers, structured output, tool use, MCP servers, embeddings, speech-to-text and web search.
@operaide/database: A SQL database per app instance with typed queries and file storage.
@operaide/vector: Embedding and retrieval pipelines for RAG, with every stage replaceable.
@operaide/document: PDF, Word, Excel and web pages turned into Markdown for your agents.

Built in

The platform does the rest.

Four that save you the most code. Tracing, usage reporting, backups and organisations come with it too.

See everything built in

REST and MCP
Every agent becomes an authenticated REST endpoint with an OpenAPI spec, and can be called via MCP. No endpoint code.
Your own interface
Ship a web interface with your app, like a review queue. Login and roles come from the platform.
Central credentials
Keys and connections are set up once by authorised people. No secrets in your code, none in logs or traces.
Roles and permissions
Platform roles plus your own permissions per app, enforced on the server. No auth checks to write.

Operaide Code

A coding agent that knows Operaide.

Our coding agent for the terminal, with the Reaktor handbook built in. Run it in Studio or on your own machine.

operaide-codeShell

1$ operaide-code2/plan3> Send invoices over 10,000 EUR to review before they are booked.4  Reading the handbook: aktor, branching5  Plan6  1. Add needsReview to the Invoice schema7  2. Branch after the extraction: over 10,000 EUR goes to review8  3. Run npm run check9/normal10> Go ahead.11  Edit src/extract-invoice.reaktor.ts (+14 −2)  Apply? [y/n]

Handbook built in: It looks up the handbook before it touches Reaktor code. The handbook ships with each release, so it always matches the version you run.
You set the pace: Plan mode only reads and proposes. Normal mode asks before every change and shows the diff. Auto mode works on its own.
Safe by design: It works only inside your app folder and runs a short list of allowed commands. Provider keys stay on the server.
Your AI models: Use the AI models your organisation has approved in Studio, or your own keys for Anthropic, OpenAI, Mistral, Azure or Ollama.

Debugging

See exactly what happened.

Every run traced step by step, with AI model, duration and tokens. No black box.

Request handling / TracingRun #2481 · 1.9 s

Trace

ReceiveRequestFunction12 ms
ClassifyAgentAgent948 ms
Classify promptPrompt8 ms
AICall1,240 tokens930 ms
DraftAgentAgent880 ms
Draft promptPrompt8 ms
AICall860 tokens860 ms
SaveCaseFunction40 ms
SendForApprovalFunction20 ms

ClassifyAgent · AI call

Model: Mistral · self-hosted
Duration: 930 ms
Tokens: 1,180 in · 60 out

Every step traced · every AI call metered

Trace walk

Step through a recorded run, or follow one live.

Structured logs

Every log line tied to its run. Filter, search, export.

Debug mode on demand

Detailed tracing for one agent, only while you investigate.

Community Edition

Free to build. Access by application.

The Reaktor framework, all packages and Operaide Studio with AI coding agents, self-hosted, free of charge.

Apply for access

Apply

A short form. We open access in stages.

Install

Self-hosted with Docker, on your machine or server.

Build

Start from a template, with any AI provider or local models.

Learn

Get started.

FAQ

Developer questions

Build your first AI app.

Apply for the Community Edition and start with a template, or bring your team and talk to us about Enterprise.

Apply for the Community Edition Talk to sales

Agentic AI as code.In TypeScript.

Less plumbing. More agents.

TypeScript your coding agent already knows

No auth, no tenants, no ops code

Your security team will say yes

A few lines of TypeScript. A production API.

Typed output

AI model per deployment

Instant API

From new app to production.

Start

Run

Publish

Building blocks for common agent tasks.

The platform does the rest.

REST and MCP

Your own interface

Central credentials

Roles and permissions

A coding agent that knows Operaide.

See exactly what happened.

Trace walk

Structured logs

Debug mode on demand

Free to build. Access by application.

Apply

Install

Build

Get started.

Documentation

Your first app

Operaide Code

Foundation Certification

GitHub

YouTube

Developer questions

Build your first AI app.