This is really cool! I’m doing something similar with Lemon Agent (https://github.com/felixbrock/lemon-agent). What I find most powerful about defining workflows in a json file is that you can add additional fields to let the LLM know about specific execution requirements, like asking the user for permission before executing a specific workflow step. This allows for infinite configuration options. Curious to hear if you already experimented with something like this or if you are planning to include something similar?
I think Zapier's NLA solution goes in a similar direction and as far as I know, they are running their model on the server side to adjust workflow input and output of workflows to the unique needs of an LLM model. They are doing an amazing job there and I am also currently working on improving this (e.g. adjusting to a model's token limit).
Besides that, Lemon Agent focuses on providing an out-of-the-box agent that is great at solving tasks with workflow automation. If you want to use Zapier NLA you have to rely on generalistic agents like LangChain's ReAct Agent or the Plan and Execution Agent or you have to build your own agent. In any case, you have to twist the agent a lot to reach a certain degree of accuracy and reliability.
I think it’s best to view this separately. Langchain provides its own agent type (next to the commonly used ReAct agent) to interact with OpenAI functions. Lemon AI functions are standalone fallback mechanisms that will be used by the model in specific situations where near-deterministic model behavior is especially important, a bit similar to the functions offered by OpenAI just suitable for open source models. You can combine Lemon functions with any Langchain agent type. Hope this helps!
Foundry is doing a nice job bringing functions in LLM apps to life. That's something I am also working on with Lemon AI functions. Besides that, to me it seems that it’s focused on read operations for web-based services. What I’m building is focused on allowing for read and write operations, both for internal tooling and web services to allow agents to take actions based on instructions. I’m essentially focussing on helping Langchain users to establish end-to-end workflow automation as quickly as possible. Going forward permission workflows will be covered as well so assistants can be team / user specific, something that I believe isn’t solved well by Zapier NLA right now.
https://simonwillison.net/2026/Feb/4/voxtral-2/
And the costs of usage are very low. For instance, with voxtral-mini-latest, you you can dictate for 5.5 hours and only have to pay $1:
https://mistral.ai/pricing#api