Why I Stopped Using JSON Tool-Calling for My Coding Agent
A few months ago I needed an agentic coding assistant inside a internal CI repair tool I was building. The obvious move was to embed something that already exists — Claude Code, or Google's Antigravity SDK. Both work well. Neither one gets you what you actually need once you're shipping a product on top of it. Claude Code isn't embeddable as infrastructure. You can shell out to it, but you don't…
A few months ago, the author needed an agentic coding assistant for an internal CI repair tool they were building. They considered using Claude Code or Google's Antigravity SDK, but found both lacking as they couldn't be fully embedded and were dependent on external infrastructure. To address this, the author developed CodePilot, an open-source, embeddable Python runtime for coding agents.
CodePilot eliminates the need for JSON tool calls and markdown, opting instead for a plain text search and replace protocol. This approach allows the model to write arbitrary Python code directly into a file named codepilot.py, which the runtime then executes, granting it more flexibility and power compared to traditional tool schemas.
The runtime also includes additional features such as a headless VT100 terminal emulator, explicit archive and reveal tools for context management, and support for multiple providers. CodePilot is MIT licensed, supports various AI models, and can be installed via pip install codepilot-ai.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
