Meatbag

Meatbag solves a specific annoyance in agent workflows: an agent walks you through generating a client secret, the instructions get buried in chat history, and now you’re not sure how to hand the secret back safely. Instead, the agent drives a shared to-do list. It creates items, nests them, and requests structured inputs; you work through them in a local web UI and check them off. The instructions live in the list, not in the scrollback.

How it works

• Agent-driven, human-completed. The agent runs meatbag from its shell to build lists and request inputs. You fill inputs, approve permission-gated steps, and mark items done in the web UI.
• Typed inputs. Items can ask for text, file uploads, secrets, or approval of a gated action, each with its own schema so the form is exactly what the step needs.
• Event-driven, no polling. meatbag wait lets the agent register listeners before it prompts, so it wakes the instant you change something rather than polling for updates.
• Local-first storage. List state is plain YAML under ~/.meatbag/, uploads are content-addressed blobs on disk, and secrets go to the macOS Keychain or 0600 files on headless Linux - the backend is chosen at build time.

Notes

It ships as a Go binary with an embedded React UI and a background daemon; make install does an atomic swap and restarts the daemon so upgrades are live. Any agent can pick it up: meatbag agent snippet prints a short markdown blurb you paste into your agent config so it knows the tool exists.