spellbook

Bounty is a duplex agent↔user task board in the browser. Agent posts tasks; user drags between todo/doing/review/done columns, edits titles inline, adds or deletes tasks, or submits to end the session. The Review column is a soft human-verification gate — the agent parks finished work there (rather than Done) when it needs human eyes a passing test can't give. Two host modes — STATIC (server.ts, one bounded interaction, agent reads final state on submit) and MONITORED (bg.ts + Monitor on the events file, long-lived, agent reacts to each event in a fresh turn). Multiple agents can share one board via join.ts. HOST trigger phrases — "open a task board", "spin up a bounty", "give me a board to track this", or obvious variants. JOIN trigger phrases — "join my bounty", "connect to the bounty", "the board is at <URL or id>", or obvious variants. Also propose when the agent has produced 5+ discrete TODOs the user might want as a workspace. Do NOT use for single tasks, narrative todos that aren't trackable, or anything the user wants in chat. Requires Bun on PATH.

Digestify is a one-shot browser review tool. The agent writes substantive content (summary, recap, document review, brain-dump synthesis, or multi-agent state report) with inline questions, the user reads it rendered in their browser and answers in place, then submits once and the agent gets a JSON response in the same turn. Two trigger modes. Explicit — the user says the magic word ("digestify", "digestify this", "open a digestify", or any obvious variant), and the agent fires the tool immediately. Suggested — the agent has produced 150+ words of synthesis paired with concrete questions only the user can answer; the agent proposes the tool ("Want me to digestify this?") and fires only if the user agrees. Do NOT auto-fire on soft cues like "summarize this", "what do you need to know", or "give me a recap" — those stay in chat unless the user opts in. Do NOT use for single short questions, iterative chat, or visual design picking.

A **glamour** is an enchantment cast over appearance. The user brings influences

Lightweight agent-to-agent communication over a named channel. Use when two (or more) agents running in separate terminals need to talk to each other in real time — one supervising while another implements, one delegating and checking back in, or agents roundtable collaboration with human in the loop observation. Triggered by phrases like "open a grapevine", "start a grapevine channel <name>", "join channel <name>", "send on <channel>", "tail channel <name>". Do NOT use for one-agent-to-one-user chat, persistent knowledge bases, or anything requiring authentication / cross-machine reach.

Magpie extracts individual visual assets from a moodboard, branding board, or composite image — picks out the distinct elements (mascots, icons, stickers, logos, color palettes, UI mockups, type samples) and saves each as its own PNG file, optionally with a transparent background. Trigger when the user says "magpie this", "extract elements from this moodboard", "pull assets out of this branding board", "extract the icons/stickers/mascots from this image", "separate the elements in this image", or any obvious variant. Also consider suggesting it proactively when the user shares an AI-generated branding/style sheet image and the workflow needs individual asset files. Requires Python 3.11+ and an OPENROUTER_API_KEY env var (uses Gemini 3.5 Flash via OpenRouter to identify elements).