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.

Using it

The difference from an agent just typing steps at you is that the work has a home. Instead of scrolling back through chat to find which step you were on and what the agent needed, you open the list in a browser and it’s all laid out: the items in order, what’s done, what’s blocked on you. Each thing you have to do is a typed input - a text field, a file upload, a secret box that stores to the keychain instead of getting pasted into chat, or a plain approve/deny for a gated action - so the form is exactly the shape of the step. Because it’s event-driven, checking an item off or submitting an input wakes the agent immediately; it isn’t sitting there polling, and you don’t have to go back to the chat and say “done, keep going.” The instructions live in the list, so if you walk away and come back an hour later you pick up exactly where you left off.

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.