llm-coding-agent

llm-coding-agent is an open-source Python plugin for Simon Willison's LLM tool. It adds an llm code command that starts an interactive coding agent session in the current directory, using any tool-capable model that LLM supports. It also exposes a Python CodingAgent class and a CodingTools toolbox.

What It Is

The project is a Claude Code-style coding agent built on top of LLM. The model can read, search and edit files under the session directory and run shell commands there. The README says the first working alpha was built from prompts given to a coding assistant, and installation currently requires the --pre flag because it depends on an LLM alpha release.

How the approval workflow works

Read-only tools run freely. File writes, edits and shell commands show what the model wants to do and ask for approval: y approves once, a approves similar actions for the rest of the session, and anything else declines, in which case the model is told and can try another approach. Options such as --yolo auto-approve every tool call, and --allow pre-approves matching commands. In-session commands include !quit, !yolo, !model MODEL and !tokens.

Tools and sandboxing

CodingTools is an llm.Toolbox confined to a root directory. It provides read_file (numbered lines, with offset and limit), write_file, edit_file (exact string replacement returning a unified diff), list_files, search_files (uses ripgrep when installed, otherwise a pure-Python fallback) and execute_command. Commands default to a 120-second timeout, capped at 600 seconds, and the whole process tree is killed on timeout with partial output returned. Paths that escape the root via .., absolute paths or symlinks return an error string instead of content.

Python API and logging

CodingAgent runs the full agent loop with a chain_limit (default 25). Its approve parameter accepts True, a callable, or None; with None the run pauses and agent.resume() continues it, which suits applications with no terminal attached. Sessions are logged to LLM's SQLite database, so llm logs shows full transcripts including tool calls, and conversations can be continued with llm code -c or --cid. The toolbox also works with llm chat --tool CodingTools.