Seahaven
An MIT-licensed open-source Python framework designed for building reproducible reinforcement learning training environments for LLM agents. It provides stateful SQLite databases, millisecond-level environment resets, parallel instances, and deterministic replay, while maintaining compatibility with the OpenEnv interface specification.
Timeline (last 90 days)
Seahaven is an MIT-licensed open-source Python framework for building reproducible RL training environments for LLM agents.
The project name Seahaven is taken from the fictional town of Seahaven in the film The Truman Show.
Every Seahaven world is an OpenEnv environment that can be directly integrated with TRL's OpenEnv support or any OpenEnv-compatible trainer.
Each Seahaven instance has its own independent SQLite database, which the agent's tool calls read from and write to.
Seahaven's Fixtures mechanism can freeze the starting state and reuse it across episodes, with reset delivering a fresh copy in milliseconds.
A single Seahaven process can accommodate hundreds of parallel instances, with each connection receiving an independent instance.
Seahaven records the before and after values of every row of data modified by the agent, enabling state-based verifiable rewards, so that most tasks do not require an LLM judge.
Seahaven is deterministic: the same fixture, clock, and random seed produce an entirely identical episode, with even random() and CURRENT_TIMESTAMP in SQL replayed exactly.
Seahaven provides both synchronous and asynchronous clients, and worlds can be published to Hugging Face.
Seahaven provides the seahaven mcp command to publish a world as an MCP server, and seahaven serve to provide a web console.
2 more timeline events
All Facts (12)
Seahaven is an MIT-licensed open-source Python framework for building reproducible RL training environments for LLM agents.
50%UnverifiedThe project name Seahaven is taken from the fictional town of Seahaven in the film The Truman Show.
50%UnverifiedEvery Seahaven world is an OpenEnv environment that can be directly integrated with TRL's OpenEnv support or any OpenEnv-compatible trainer.
50%UnverifiedEach Seahaven instance has its own independent SQLite database, which the agent's tool calls read from and write to.
50%UnverifiedSeahaven's Fixtures mechanism can freeze the starting state and reuse it across episodes, with reset delivering a fresh copy in milliseconds.
50%UnverifiedA single Seahaven process can accommodate hundreds of parallel instances, with each connection receiving an independent instance.
50%UnverifiedSeahaven records the before and after values of every row of data modified by the agent, enabling state-based verifiable rewards, so that most tasks do not require an LLM judge.
50%UnverifiedSeahaven is deterministic: the same fixture, clock, and random seed produce an entirely identical episode, with even random() and CURRENT_TIMESTAMP in SQL replayed exactly.
50%UnverifiedSeahaven provides both synchronous and asynchronous clients, and worlds can be published to Hugging Face.
50%UnverifiedSeahaven provides the seahaven mcp command to publish a world as an MCP server, and seahaven serve to provide a web console.
50%UnverifiedSeahaven runs entirely locally, depends only on Python 3.14+ and SQLite, and requires no external services.
50%UnverifiedSeahaven was created by the Kiln team and is used internally for evaluating and optimizing agents, but it is an independent project that can be used without Kiln.
50%