Genkit
Genkit is Google's open-source framework, not a hosted service, for building full-stack, AI-powered and agentic applications. One unified interface reaches many model providers, with SDKs in TypeScript, Go, Python and Dart, local tooling, and no charge for the framework.
What is Genkit?
Genkit is Google's open-source framework for building full-stack, AI-powered and agentic applications, presented on its overview page as built and used in production by Google. It is a framework, not a hosted service: you install it into your own project and run it on your own infrastructure.
Four SDKs share a coherent API surface—TypeScript for Node.js, Go, and Python and Dart, both still in Preview. The project rests on five stated pillars: model-agnosticism, platform portability, rich local tooling, complete observability, and seamless integration into user-facing applications. Its roadmap describes Genkit as having graduated into Google's AI framework at company scale.
The central abstraction is a unified generate() call. The same function reaches Gemini, GPT, Claude, Grok, DeepSeek, Qwen, Kimi, Z.ai or a local Ollama model; switching provider means changing a plugin and a model identifier, not rewriting the application. Around it sit the framework's primitives: flows (typed, streamable, traceable functions), Dotprompt templates, tool calling, runtime context, middleware, interrupts, persistent chat, RAG and Model Context Protocol support. A beta Agents API adds chat(), sendStream(), loadChat() and remoteAgent(), together with session stores, background execution and multi-agent delegation.
Local tooling is a deliberate strength. The genkit-cli command line tool, installed through npm or as a native binary, launches a Developer UI on localhost:4000 with a telemetry API on port 4033. That UI exposes action runners for flows, prompts, models, tools, retrievers, indexers, embedders and evaluators, so each step can be run, compared and inspected before shipping. In production, Genkit Monitoring reports latency, errors and token usage and exports traces you can replay as evaluations, routed through the Firebase plugin and the Cloud Logging, Trace and Monitoring APIs.
Deployment targets include Cloud Functions for Firebase, Google Cloud Run, Azure Functions, AWS Lambda or any Node.js host, explicitly with or without Google services. More than twenty web and back-end frameworks are documented, from Next.js and Angular to Django, FastAPI and Gin, alongside roughly a dozen vector stores. The core has been generally available since version 1.0 under semantic versioning, with unstable work isolated in a separate genkit/beta channel. The audience is development teams that want provider independence and production-grade observability without adopting a hosted platform.
What it does
- Call hundreds of models from many providers through a single generate() API
- Orchestrate AI workflows as composable, typed flows
- Produce structured, schema-validated output and let models call your own tools
- Build conversational agents with persistent history, sessions and streaming
- Run retrieval-augmented generation over a vector store, then generate
- Trace, debug and evaluate your AI logic locally in the Developer UI
- Monitor latency, errors and token consumption once in production
When to use Genkit / When not to
A quick filter to help you decide if Genkit is the right fit.
When to use Genkit
- Back-end and full-stack developers adding an AI feature to an application they already run
- Multi-language engineering teams working across TypeScript, Go, Python and Dart with one consistent API
- Teams that want to stay model-agnostic and swap provider without rewriting their application
- Firebase and Google Cloud Run users, who get onCallGenkit integration and Genkit Monitoring out of the box
- MLOps and LLMOps engineers who need traces, metrics and evaluation once AI logic reaches production
When not to use Genkit
- Non-technical users looking for a no-code interface: everything runs through code written in one of four languages
- Anyone expecting a ready-to-use consumer application, since Genkit only exists inside a development project
- Teams without a model provider account: Genkit ships no models and no API keys of its own
- Organisations that require published terms of service, a privacy policy or a DPA before adoption, none of which exists on the domain
- Developers working in Java, C#, Ruby, PHP or Rust, as only TypeScript/JavaScript, Go, Python and Dart are covered
How to use Genkit
A typical end-to-end flow, from setup to results.
- Choose the SDK for your language: TypeScript, Go, Python (Preview) or Dart (Preview)
- Pick a model provider and obtain an API key; Vertex AI uses a different authentication method
- Install the Genkit SDK, the provider package and the CLI (npm install -g genkit-cli, or the native binary via curl -sL cli.genkit.dev | bash, or the Windows executable)
- Initialise Genkit in your configuration file with the provider plugin loaded
- Write your AI logic: prompts, flows and tools, with typed input and output schemas
- Run genkit start followed by the command that executes your code, to open the Developer UI on localhost:4000
- Test and iterate in the Developer UI: run flows and prompts, compare outputs across models, inspect traces step by step
- Use the CLI for repeatable runs: genkit flow:run, genkit eval:flow, genkit trace:list, genkit trace:get
- Expose your flows as HTTP endpoints through your back-end framework, or wire up a frontend with the client helpers
- Deploy to Firebase, Cloud Run, Azure Functions, AWS Lambda or any Node.js host, then enable Genkit Monitoring (ENABLE_FIREBASE_MONITORING=true or enableFirebaseTelemetry())
Pros & Cons
Pros
- Free and open source, with no paid tier and no quota imposed by the framework itself
- Genuinely model-agnostic: changing provider does not mean rewriting the application
- Four languages with consistent APIs, TypeScript, Go, Python and Dart, which is rare among AI frameworks
- Outstanding local tooling: Developer UI, step-by-step traces and side-by-side model comparison
- Deployment portability claimed explicitly, with or without Google services
- Backed by Google and used in production by Google, generally available since 1.0 under semantic versioning
- Production observability designed in from the start, plus a broad ecosystem of frameworks, vector stores and MCP support
Cons
- No legal documentation whatsoever on the domain: no terms of service, no privacy policy, no DPA, no sub-processor list and no GDPR mention
- Developers only: nothing can be used without writing code, in one of four supported languages
- No named support channel, neither email address nor contact form, only Discord, GitHub Issues and the genkit tag on Stack Overflow
- Python and Dart are still in Preview, while the Agents API, Chat/Sessions and Interrupts are in beta and can break on a minor release
- Genkit Monitoring requires a Firebase project on the paid Blaze plan plus usage-billed Google Cloud services, so production observability is not free
- The real cost is displaced rather than removed: model provider keys, hosting and observability are all billed elsewhere
- Strong pull towards Google and Firebase despite the portability promise, since monitoring runs through the Firebase plugin
Pricing & Plans
Genkit is distributed as an open-source framework and is free of charge. The site publishes no pricing page—a scan of the 299 URLs in its sitemap confirms it—and states no amount, no currency and no billing unit; the overview page notes that its sample applications may be explored at no cost to you. The relevant costs are therefore indirect and fall entirely to the developer: API keys and per-call charges from the chosen model providers, hosting for the deployed back end, and, for Genkit Monitoring, a Firebase project on the paid Blaze plan together with Cloud Logging, Trace and Monitoring usage billed under Google Cloud Observability rates. Prospective users should budget for those third-party charges rather than for the framework itself.
Data, GDPR & hosting
A consolidated view of how Genkit handles your data.
GDPR overview
There is no GDPR mention anywhere on genkit.dev. The term appears on none of the sixteen pages reviewed, in text or in HTML. The domain publishes no privacy policy, no trust or security page, no data processing agreement and no sub-processor list. No Article 27 representative is designated, no data protection officer is named and no GDPR contact address is given. The only privacy-related reference on the entire site points to a Google document, the cookie policy at policies.google.com/technologies/cookies, linked from the home page cookie banner and from the Analytics section of the developer tools page. The practical consequence is that Genkit does not address compliance at all: as a library running inside your own infrastructure, it leaves you the controller for your application, and any assessment has to be made against your chosen model provider and hosting arrangements.
Who owns the data?
genkit.dev publishes no terms of service, no privacy policy and no legal notice—a scan of the 299 URLs in its sitemap confirms it—so there is no ownership clause to report, neither a waiver nor a guarantee. The question moves elsewhere. Genkit is a library installed in your own project, so application data never passes through Genkit as a service: it reaches whichever model provider you configure (Google, OpenAI, Anthropic, xAI, DeepSeek and others) under that provider's terms, or stays local with Ollama. The only collection attributable to Genkit is CLI and Developer UI analytics, which the documentation describes as cookies and similar technologies from Google. Genkit Monitoring writes traces into your own Google Cloud project.
Reuse rights
No reuse clause is published on genkit.dev, because the domain carries no terms of service at all. What the architecture implies is nonetheless clear: the code you write with Genkit and the outputs your models return stay in your own project, since Genkit never interposes itself between your application and the provider. What you may do with those outputs is therefore governed by the terms of the model provider you call—Google, OpenAI, Anthropic, Ollama and the rest—and not by Genkit. The project describes itself as open source in its title, meta description, home page, overview and roadmap, but the licence text is not published on genkit.dev; it lives in the genkit-ai/genkit repository on GitHub, alongside the CONTRIBUTING.md and CODE_OF_CONDUCT.md files that govern contributions. Read both the provider's terms and the repository licence before assuming any right of reuse.
Data retention & training
Hosting summary
Genkit announces no hosting jurisdiction, for a structural reason: it hosts nothing. The framework is a library that runs inside your own project, so there is no Genkit-operated storage layer to locate, and where your data physically sits is entirely a consequence of your own choices. Your back end runs wherever you deploy it: Cloud Functions for Firebase, Google Cloud Run, Azure Functions, AWS Lambda or any Node.js host. Model calls leave for the provider's infrastructure (Google AI Studio, Vertex AI, OpenAI, Anthropic, xAI, DeepSeek, DashScope, Kimi, Z.ai, OpenRouter, Azure or AWS Bedrock), or never leave your machine at all when you run a local model through Ollama. Vector stores are likewise selected and hosted by you. If you enable Genkit Monitoring, traces and logs are written to Google Cloud Logging, Trace and Monitoring inside your own Google Cloud project. The documentation site itself is served through Fastly (AS54113, anycast address 199.36.158.100, Mountain View node in the United States), but that concerns the site, not user data. No sub-processor list is published.
Things to keep in mind
Risks and trade-offs to weigh before adopting Genkit.
- No privacy policy or terms of service on the domain, so you have no contractual statement of what is collected or on what basis
- Application data leaves for the model provider you choose, whose terms are neither summarised nor linked from genkit.dev
- Genkit Monitoring exports traces, inputs and outputs to Google Cloud, where sensitive data can end up if your code does not filter it
- CLI and Developer UI analytics are on by default; the opt-out has to be set by hand with genkit config set analyticsOptOut true
- Indirect costs are easy to underestimate: usage-billed model calls, hosting, the Firebase Blaze plan and Cloud Logging, Trace and Monitoring
- Autonomous agents with tool calling produce real side effects; interrupts and tool approval are documented, but they only help if you actually use them
- Delegating orchestration to a framework can erode a team's grasp of the model loop, and outputs still need human verification: Genkit traces and evaluates, it does not judge correctness for you
Setup & Integrations
Technical difficulty
The audience is exclusively developers: nothing works without code. Prerequisites are modest: a working environment for your chosen language, an API key from a model provider and a package install. The first steps are easy: the site promises a start in just a few lines of code, each getting-started guide is self-contained from an empty project to a running app, and the CLI installs through npm or a native binary. Difficulty rises later, with deployment, authentication, session stores, Vertex AI's non-key authentication, and Genkit Monitoring, which requires a Firebase Blaze project, three Google Cloud APIs and three IAM roles.
Deployment
Integrations
Supported languages
Behind Genkit
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
What is Genkit?
Which programming languages does it support?
Is Genkit free?
Which models can I call?
Do I have to use Google Cloud?
Is Genkit ready for production?
How do I test and debug my AI logic?
Where can I get help?
Does Genkit publish a privacy policy or terms of service?
Is my data used to train models?
Should you pick Genkit?
Genkit occupies a clear position: an abstraction layer over many model providers plus the tooling to develop against them, and not a hosted service. Its strongest argument is consistency—the same primitives, the same generate() call and the same Developer UI across TypeScript, Go, Python and Dart—matched by local tooling few competing frameworks reach. Credibility follows from its origin: built by Google, used in production by Google, generally available since 1.0 under semantic versioning.
Two reservations deserve weight. The first is the complete legal silence on the domain: no terms of service, no privacy policy, no data processing agreement and no GDPR mention across 299 indexed URLs. That is unusual for a tool entering an enterprise stack, and it means the governing terms are always someone else's, those of the model provider you call and of the infrastructure you deploy on. The second is uneven maturity: Python and Dart remain in Preview, and the Agents, Chat/Sessions and Interrupts APIs are in beta and may break on a minor release.
Cost deserves the same clarity. The framework is free, but production observability is not: Genkit Monitoring depends on a Firebase project on the paid Blaze plan and on usage-billed Google Cloud services, while model calls and hosting are billed separately.
The 2026 roadmap points to Dart reaching 1.0, Kotlin and Swift client SDKs, an Agent Primitive, middleware, A2A work towards the Gemini Enterprise Agent Platform, and the monorepo splitting into per-language repositories.
For development teams that want provider independence with serious observability, Genkit is a strong choice. Before committing, read the terms of the model provider you intend to use: they, not Genkit, are what actually govern your data.
- Choosing a selection results in a full page refresh.
- Opens in a new window.