# Koog
Koog is [[JetBrains]]' open source framework for building [[AI Agents]] on the JVM. Kotlin-first, with a proper Java API next to it. Apache 2.0. Version 1.0 shipped in May 2026 with a promise that matters more than any feature in the release notes: no breaking changes to stable modules for at least a year.
Most agent frameworks assume Python. If your business logic already lives in Kotlin or Java, that assumption costs you a service boundary, a second deployment, and a second set of on-call habits. Koog exists so you don't pay that tax.
## What it gives you
- **Agent strategies as graphs** — nodes, edges and subgraphs instead of one opaque loop. You can read the control flow.
- **Tools** — typed tool definitions, parallel execution, plus [[Model Context Protocol (MCP)]] servers as tool sources. It also speaks [[Agent Client Protocol (ACP)]], so agents can plug into standard clients.
- **Provider portability** — Google, [[OpenAI]], [[Anthropic]], DeepSeek, OpenRouter, [[Ollama]], Bedrock. Switch provider mid-conversation without losing history.
- **History compression** — automatic context trimming for long-running runs, so token cost doesn't grow without bound. Anthropic prompt caching is wired in too.
- **Persistence and memory** — checkpoints you can restore from, plus longer-lived agent [[AI Agent Memory|memory]]. Retries are built in rather than bolted on.
- **Observability** — OpenTelemetry across all targets, with exporters for Langfuse and W&B Weave.
- **Framework integration** — Spring Boot and Ktor, so an agent is just another bean or route.
## Where it runs
JVM, Android, iOS, JS, and WasmJS through Kotlin Multiplatform. The 1.0 release added local Android inference via LiteRT, which means an on-device agent with no network round trip. Requires JDK 17+ and Kotlin 2.3.10+.
Install is one line: `implementation("ai.koog:koog-agents:1.1.1")`.
## Why I'm paying attention
Two reasons.
First, the stability commitment. Agent frameworks churn violently right now, and a year of API stability from a company that ships IDEs for a living is worth more than a longer feature list. You can build on it.
Second, the interop story. Decoupled HTTP transport, a redesigned Java layer, MCP and ACP support: Koog is built to sit inside an existing system rather than become the system. That's the right shape for enterprise work, and it's where [[LangChain]] and friends often struggle.
I don't write Kotlin daily. But if you do, this is the option I'd start with instead of shelling out to a Python sidecar.
## References
- Product page: https://www.jetbrains.com/koog/
- Documentation: https://docs.koog.ai/
- API reference: https://api.koog.ai/
- Repository: https://github.com/JetBrains/koog
- Releases: https://github.com/JetBrains/koog/releases
- Maven Central: https://central.sonatype.com/artifact/ai.koog/koog-agents
- Issue tracker: https://youtrack.jetbrains.com/issues/KG
- Koog 1.0 announcement: https://blog.jetbrains.com/ai/2026/05/koog-1-0-is-out-stable-core-better-interop-and-multiplatform-observability/
- The Kotlin AI Stack: https://blog.jetbrains.com/kotlin/2025/09/the-kotlin-ai-stack-build-ai-agents-with-koog-code-smarter-with-junie-and-more/
- Kotlin AI app development overview: https://kotlinlang.org/docs/kotlin-ai-apps-development-overview.html
## Related
- [[JetBrains]]
- [[AI Agents]]
- [[AI Agent Memory]]
- [[AI Agent Orchestration]]
- [[Agentic Engineering]]
- [[Model Context Protocol (MCP)]]
- [[Agent Client Protocol (ACP)]]
- [[LangChain]]
- [[OpenAI Agents SDK]]
- [[Pydantic AI]]