true-code

true-code is an agentic coding harness that shows what it did and what it cost. It is an open-source tool that reads actual code before answering, using tools such as read_file, list_dir, glob, grep, and in write mode write_file and patch, and in full mode shell. Every call is visible as it happens. Nothing changes without the user seeing the diff: every write is previewed and confirmed before it is applied, and every command is shown before it runs. There are three permission modes, defaulting to the safe one, and a tool the mode does not allow is never offered to the model at all. Guard rails stop with a reason—turn limit, session budget, and loop detection—and live token and cost accounting includes a budget that halts the run. The harness provides evidence instead of "done": every run ends with what the harness observed, including files changed, commands run, and exit codes. Changed code that nothing checked is reported as UNVERIFIED in red. Project rules written in .truecode/constraints.toml are restated on every request and verified against the change before it is applied. Undo is available via /undo in the TUI or truecode undo from the shell, reading the session log. An append-only session log in .truecode/sessions/ records what ran, what was allowed, and what can still be taken back. It supports three providers—Anthropic, OpenAI, and OpenRouter—with tool calls and prompt-cache accounting on every one. Headless mode (-p) is available for scripts and CI. Not yet built: OS sandboxing, MCP, and the LEARN mode.