Give the AI a study, not your whole phone
Foreverse ships an in-app agent that does real work: your novels, character cards, and lorebooks are files it can read and write. Task logs from development: a 121-round chained treasure hunt (60/60 found), growing a lorebook out of a novel, forking a card from the library and starting a chat, shopping the community on your behalf — and the moment a failed edit fed its error back and the agent fixed itself.
The phrase “AI agent” has been stretched until it barely means anything, so this post skips the concept and prints the logs. One line of context: the Agent tab in Foreverse hosts an agent with four tools (read, write, edit, execute), and every novel, character card, and lorebook on your shelf is a file in its workspace. Below are five things we made it do during development and testing, failures included.
Task one: a 121-round treasure hunt, 60 of 60 found
This is the stress test for the tool-calling loop: hide 60 “gem” files in the workspace, each holding a clue to the next, some clues locked until a file is edited correctly. Running on DeepSeek's official API, the agent chained 121 tool calls and found all 60 gems with zero human intervention.
More valuable than the clean sweep was the stumble around round 40: an edit missed its target text by a few characters, and the tool layer fed the failure message straight back. The agent read the error, re-read the file, and landed the edit on the second try. Self-correction after error feedback is the dividing line between an agent that can work unattended and one that cannot; that exact scenario is now a regression test.
Task two: growing a lorebook out of a novel
Give it an imported long novel and the instruction “organize the setting into a lorebook.” Its approach resembles a careful human editor's: read the text in slices, note characters, places, and factions as they appear, then write entries into the book's world directory in proper lorebook format, keywords and aliases included. The finished entries feed directly into the same lorebook runtime the reader and tavern chat use — triggered by keyword during continuation and roleplay.
The cost structure deserves a sentence: reading a 300k-word book is not one giant prompt but dozens of read-a-slice, note-a-slice rounds. Per-round context stays small, so a budget model handles it fine.
Task three: fork a card, then start the chat
We ran a blind test once: no hints, just “make a new card in the style of that one in the library.” It read the character library on its own, found the reference card, wrote a new one with the same field structure, and passed format validation on the first try. When it finished, the chat screen offered a start-conversation card; one tap and the tavern session opened.
A product stance we hold onto here: whatever the agent creates must land as standard-format files — chara_card_v3, lorebook JSON — not rows in a proprietary database. A card it writes can be exported, shared, and carried to desktop SillyTavern.
Task four: shopping the community for you
“Find me a cyberpunk card, check that it looks legit, install it if so.” The agent can browse the community, read detail pages, and download-and-import on your behalf, with a dedicated confirmation card before any download. Hard rules apply: browsing and detail reads are confirmation-free (read-only); downloads need your approval or a pre-enabled auto mode; each turn carries read and download quotas; and everything installed keeps a provenance record — which community pack, which version — written into its metadata. It is not an unlimited scraper; it is a personal shopper with a budget and receipts.
Task five: errands without leaving the page
The reader, tavern, and companion screens carry a floating dock. Tap it and a mini overlay opens: streaming bubbles, compact approval cards, @-references to the current book or card. Halfway through a chapter you can say “file this foreshadowing into my notes” without leaving the page. Tasks run in the background; switch screens or lock the phone, and the in-flight reasoning stream and tool cards are still there when you return.
Boundaries, stated plainly
Two design boundaries. First, no device permissions: no contacts, no SMS, no screen automation. Projects like RikkaHub's agent fork explore that direction with 80-plus device tools; it is a legitimate road that is not ours. Full power inside the study, none outside it — every permission beyond the library is trust you should not have to extend. Second, it spends your tokens: long tasks are real API consumption, so runs show a visible step count and a stop button that always works, and whatever finished before the stop is kept.
The full guide to the agent and its skill catalog lives in the docs; the product page is here.
FAQ
How is this different from a chatbot?
It has tools. Read, write, edit, and execute run over the model's native tool-calling protocol, and every book, character card, and lorebook on your shelf is a real file in its workspace. Ask it to 'turn this novel's cast into dossiers' and it actually reads the chapters and actually writes the entries into the world's directory — every step shows as a visible tool card, and writes wait for approval.
Can it touch anything else on my phone?
No, deliberately. Its workspace is your study: novels, cards, lorebooks, notes. Contacts, SMS, and screen automation were never granted — community projects do explore that direction (RikkaHub's agent fork ships 80+ device tools), and we chose the opposite boundary: full power inside the library, no power outside it.
Will it mangle or delete my files?
Write operations go through an approval card by default: which file, changed how, laid out on screen before anything lands on disk. Approval has three modes (ask every time / ask for risky ops only / full auto), switchable per session. Runs keep a complete reviewable log, and anything shaped like 'token: value' in tool output is masked before it reaches the model's context.
What does a task cost?
It runs on your own API key at provider prices. A 'fill in the lorebook' task is a read-write loop in the tens of thousands of tokens — pocket change on a flash-tier model. The 121-round stress test ran on DeepSeek's official API and cost less than a soft drink. You can switch models per task: flagship for hard jobs, cheap for chores.
Questions or ideas? Join our Discord →