Aviary

What it does

Aviary takes in task definitions, tools, and environment state written in Python and provides a gymnasium-style interface in which a language agent receives observation messages and returns tool-request messages. Each step returns observations, a reward, and done or truncated flags, so developers can train or evaluate agents on tasks such as scientific literature search.

People use Aviary by installing the fhaviary package and either subclassing Environment to build a custom environment or running included ones for math, general knowledge, biological sequences, literature search, and protein stability. It is designed to work with the sister library LDP, which defines agents as Language Decision Processes. The repository also lists tutorial notebooks and a Jupyter notebook environment.

Aviary suits developers and researchers who work in Python and want to build or evaluate tool-using language agents. Installation is via pip and the code is published under the Apache-2.0 license. Running agents depends on external language model endpoints the user supplies, and the included environments are narrow task-specific tools rather than a general scientific capability.

What it can do

• Define custom language agent reinforcement-learning environments with reset and step methods.
• Exchange OpenAI-style messages, including tool request and tool response messages with image content.
• Expose Python functions as agent tools, with hidden state and tool-call ID injection.
• Run included environments for math, general knowledge, biological sequences, literature search, and protein stability.
• Evaluate language agents through the sister library LDP using RolloutManager trajectory sampling.

Works with

• LDP (Language Decision Process agent library)

Before you choose

• Running agents requires a language model endpoint and any associated costs, which are separate from the Apache-2.0 library.
• Included environments are task-specific and are described as not generalizing between tasks.
• Aviary is a code library installed locally, not a hosted application.

Listed under

• Coding
• MCP and agent tools

Get started

Install with pip install fhaviary; tutorial notebooks and developer setup are described in the repository.

1. Install the library with pip install fhaviary, or with extras such as gsm8k, hotpotqa, labbench, lfrqa, and notebook for included environments.
2. Install the dev extra to run the tutorial notebooks.
3. Define an environment by subclassing Environment or using the simple environment pattern, then evaluate an agent with LDP's RolloutManager.

Aviary is released under the Apache-2.0 license. Model endpoints, compute, and dependency costs are separate and must be verified by the user.

Aviary is free to use. See their pricing page ↗

Sources for this overview

Product pages, documentation, and pricing checked 29 Oct 2026. Each source supports part of the overview; availability and prices can change.

• GitHub - Future-House/aviary: A language agent gym with challenging scientific tasks · GitHub ↗github.com
• Aviary: training language agents on challenging tasks | FutureHouse ↗www.futurehouse.org
• FutureHouse ↗www.futurehouse.org
• What is an AI Scientist? | FutureHouse ↗www.futurehouse.org

Company website

github.com: product, pricing and policy pages

Social accounts

Accounts the company links, and its posts on X

Human review

Checked by a person

knowledge_base

manual_review

Last checked 29 Oct 2026. Every field keeps its source and date; unknown means we couldn’t confirm it.