Plan in context
shad surveys the live environment and turns a plain-language goal into an explicit route through the work.
01 / 02Agent infrastructure
shad
A terminal agent that works in the shell you already use.
Give shad an outcome in plain language. It opens a real shell, plans the work, runs commands, handles interactive prompts, and reports what it found—without taking control away from you.
$ shad update my project dependencies safely
01 Reading package constraints and current lockfile
02 Checking release notes for breaking changes
shad I found one major upgrade. Holding it back; applying 7 compatible updates.
› run the full test suite before committing
✓ 84 tests passed · worktree clean
type to steerstop to abortctrl-c for shell
The operating loop
Intent in. Evidence back.
Operate the real shell
A live PTY lets it use ordinary commands, respond to prompts, follow long-running jobs, and work with the terminal state already in front of you.
Remain steerable
Type a correction while the task is running, send stop to abort, or return directly to the shell with Ctrl-C.
Remember the room
Per-terminal memory keeps follow-up work connected without turning every new task into a cold start.
Design principle
Autonomy needs an interrupt.
shad treats mid-flight steering as part of the core interaction—not an escape hatch. The human can redirect the plan at the moment new information matters.
Availability Private alpha. A public distribution channel is in preparation.