Proposes the Oats Protocol, an open architecture pattern for standardized tool-calling in local coding agents, addressin

The Oats Protocol – Open Agent Tools for Local Coding Agents Recently I was using functiongemma and watched it load and run local source code as a tool call without any training/tuning. A couple days later I got Qwen35 in Open-WebUI to use the "native" tool-calling. With Open-WebUI I could observe the changes as it ran inside the docker containers crawling over stuff on its own, but it was not obvious to observe functiongemma.

As a control freak, the differences in how these two tool-calling approaches got me thinking:

How will open source enable standardized tool-calling for agents so we do not have to build and support custom tool-calling harnesses on our own?

I wanted to share an architecture design pattern we're using to mitigate custom code for tool-calling in many components/subsystems. We open sourced our local OATs coding agent on GitHub https://github.com/district-solutions/open-agent-tools-coder. I run coder with a large local model that delegates tool calling to smaller local models. The coder includes vLLM deployments in the stacks dir https://github.com/district-solutions/open-agent-tools-coder/tree/main/stack for running Qwen36 27B and 35B with tool-calling delegation to functiongemma.

On startup, coder looks for a preprocessed, large JSON index of supported tools. We open sourced the OATs Tool-Calling Prompt Index for >141K Tools on GitHub https://github.com/district-solutions/open-agent-tools#openagent-tools-oa