A nest of a bird of prey perched high on a cliff or tree top.
Aerie is a tool for building and running AI-powered workflows. Rather than giving a language model free rein over a complex task, Aerie lets you break tasks into discrete, well-defined steps — with AI playing a focused role in key steps. The result is more predictable and debuggable than purely agent-driven systems.
Download the AppImage from the releases page:
chmod +rx aerie-x86_64.AppImage
./aerie-x86_64.AppImageThe AppImage also runs on Windows via WSL. For Nix or source builds, see the installation guide and development guide.
Workflows are structured as node graphs: each node can represent an agent, data transformation, decision, or other action. Data flows in one direction along wires connecting nodes, making execution order explicit and the overall logic easy to follow and explain.
Each node executes at most once per run. The editor also supports incremental runs, which re-execute only the nodes that have changed since the last run. This makes iteration fast during development — tweak a prompt or swap a model and only the affected portion of the graph reruns, leaving expensive upstream steps untouched.
Nodes can be grouped into subgraphs, which appear as a single node in the parent workflow, keeping complex graphs organized and readable. Subgraphs can themselves contain subgraphs, allowing complex workflows to be built up from well-defined, reusable pieces. A special iterative variant can apply a subgraph across an entire list of inputs — see Iteration below.
See the first steps tutorial 1 for an introduction to building workflows, and the subgraphs and iteration tutorial 2 for a deeper look.
Note
Tutorials are currently in progress. Links will be updated as they are published
Furthermore, workflows can dispatch to other workflows in a sequential chain, enabling dynamic routing and tail recursion.
Aerie's Structured Output node instructs an LLM to produce output conforming to a JSON Schema, making it straightforward to extract structured data from natural language or generate it from scratch. Once you have structured data, transformation and templating nodes let you reshape and render it without touching the LLM again — keeping things cheap, fast, and precise. See the structured generation tutorial 3.
Agents can interact with external services through the Model Context Protocol. Aerie manages MCP tool providers from a dedicated Tools tab, supporting both local STDIO servers and remote HTTP services. Tools can be selected per-agent, so each step in a workflow only has access to what it needs. The agent tools tutorial 4 walks through a practical example using live weather data.
You can also use the Invoke Tool node to manually make tool calls — bypassing LLM tool selection. This can be useful in cases where the arguments are known ahead of time or need to be tightly controlled. It also allows for manipulating the tool results before sending them to a language model. This is covered in the tool invocation tutorial 5.
Iterative subgraphs apply a nested workflow to every item in a list, collecting the results into an output list. This makes it possible to process inputs rigorously at scale — for instance, checking each extracted claim in a document individually rather than asking the model to handle them all at once. Optional parallelism can speed things up, though rate limiting is advisable with remote APIs. See the subgraphs and iteration tutorial 2.
Workflows don't have to be simple chat agents. Named Output nodes emit results that can be consumed by other applications — written to the console or individual files when run via the runner CLI:
aerie-runner \
--workflows ~/.local/share/aerie/workflows/ \
--model openrouter/openrouter/free \
-I article.txt \
exec my-workflowSee the [outputs tutorial] for details.
Coming soon...
Workflows can accept image inputs along with text to handle tasks like visual question answering or converting images into structured data.
Consult the documentation for detailed instructions.
Import and experiment with the Example Workflows.
The [tutorial series] builds up from a simple two-agent chat workflow through structured generation, tool use, document analysis, iteration, and batch output — each article building on the last. Start with First Steps1.
Aerie is still under active development. While it may not be ready for production use, it is well-suited for exploration and prototyping.
All projects in this repository are licensed under the Mozilla Public License Version 2.0