PurrCode

About

PurrCode starts from the position that model output is a proposal and never authority: repository content, downloaded skills, and generated code are all untrusted until a separate authorization layer approves each action, re-checks that approval immediately before execution, and records the validation afterward. Permission modes are daemon-enforced constraints rather than prompt-level politeness, so a read-only mode genuinely prevents writes, and sandboxing via sandbox-exec or Bubblewrap is reported honestly rather than oversold. Agent work runs in detached Git worktrees so uncommitted user work is never silently stashed or overwritten. A pure-Rust desktop IDE built on egui shares one daemon-owned session with the terminal TUI, letting a task move between interfaces without losing state. Local-first operation connects to Ollama or LM Studio by default with NVIDIA NIM as a first-class cloud option, and credentials stay in the OS keychain, never in model context. Developers who want a security-enforced local agent rather than a prompt-level one use PurrCode.