Coding agents
Run Claude Code, Codex, or a local Pi subprocess against your codebase with real file tools, a transcript you can read, and undo a few clicks away.
What you wire up
A coding session in Froots is just chat with the right backend and the right skills enabled. Claude Code and Codex both run as one-shot subprocesses; Pi runs as a long-lived process that keeps multi-turn memory between prompts.
The agent gets a workspace-scoped set of file tools: agent_read_doc, agent_write_doc, and agent_edit_doc (strict one-match substring replace, so the model can’t make ambiguous edits). Everything is bounded to your workspace folder.
Skills that pair well
- `github` — read, comment, and ship PRs via the
ghCLI. - `coding-agent` — Froots’s shipped guidance for code-writing flows.
- `safe-file-deletion` — always-on; gates destructive ops behind a confirm step.
- `tmux` — drive a terminal session if you want long-running processes outside the agent loop.
What it looks like in practice
Pair it with the Careful permission policy and Froots will prompt before any edit. Pair it with Yolo and it just rips.
What it doesn’t replace
Froots isn’t an IDE. It runs alongside your editor, owns the agent loop, and writes to disk; you commit the diffs in git on your own terms. Reasoning/thinking event surfacing for Claude/Codex is still in flight.