YOUR PROFESSIONAL WRITING WORKSTATION

From a single spark, a novel.

Weaver — an agent built for writing, and nothing else.

AI assists your writing. You remain the author.

If you can't bring yourself to read a story, what is it for?

Bulk book-cloningOne-click rewritingContent farming

Four doors to the same writing desk

Weaver, your dedicated editorial team

Agents built and tuned for writing, with carefully designed pipelines that keep context and tools uncontaminated. From project setup to finished chapters, all it takes is conversation.

The cockpit, under your hand

Already used to polishing prose or drafting beat sheets with AI? Manage the AI writing pipeline by hand, with every step's output under your control.

The study, longhand only

Don't want AI touching your writing, but still need organizing, feedback, and reference management? This mode offers pure writing plus easy lookup and organizing.

Polish, built for fiction

Drafted elsewhere, but it reads like AI and not like you? A polishing mode built for fiction, tuned for most mainstream models.

The real interface, unretouched

An agent pipeline built for writing, not code: conversation alone dispatches and orchestrates it.

The writing cockpit: the author's brief passes through verbatim on the left; the working copy shows a fresh English draft, worldbook retrieval, thinking steps, and De-AI and Style polish actions with per-segment cost — The cockpit: brief in, draft out, cost on the spot

Whole-chapter view: Chapter 1 assembled read-only at 483 words, with find-and-replace and .txt download — The whole chapter, ready to read

Real interface capture: one English writer run on a demo project, unedited.

Hand it the messy pile

Conversation is the whole way in. Upload three files of scribbled notes and it reads them, rules on each one, and files them into a single worldbook — saying so out loud wherever the notes contradict each other. Every move comes back as a card that names what changed, and clearing it is your call. Stuck on the plot? Bring the editor into the same window.

Weaver in conversation: the author tells it to file the notes in and make the calls itself, flagging anything it changed; tool chips read Check worldbook and Read asset; it then rules on three contradictions — tower height 96 ft, flash characteristic 12 seconds, Maren's age 19 now and 18 the night Ansel went — and files four proposals for worldbook, style, book outline and key memory, each confirmed as a saved asset with its character count — Three files of scribbled notes · read, ruled on, filed · contradictions settled on the spot

The editor thread: the author asks whether ending chapter one on both reveals at once is one reveal too many, and which to give up if either; a tool chip reads Handing over to the story editor; the editor answers that the double reveal stays exactly where it is, argues it from the author's own sentences, and closes with lock it in and keep moving; a banner explains that messages go straight to the editor while Weaver only passes along background material, never rewriting or relaying — Switch to the editor thread · one concrete question · answered out of your own lines

Both captures are real runs, unretouched, shot separately in each language (demo projects: The Lamplighter / 守灯人).

Does it read like AI wrote it?

Our pipeline turns the plot into what you meant, not a trope; and the prose into your voice, not the model's.

Every hand edit you make is remembered and avoided in the next chapter.

No more pushing up glasses that aren't there. Not “he hugged her”, but a grip that folds you into his bones, possessive to the last inch, a meteor dropped into the lake of your heart.
What de-AI'd prose should feel like

Quick De-AI result page: the raw draft from the section above after patch-style edits, 5 redundancies removed, deepseek-v4-pro, $0.0005 for this run, with a per-edit change list — The raw draft from above, de-AI'd · 5 edits · $0.0005 · Chinese edition shown

Your voice, not a template's

Style is the knob that matters most here. Start from a public-domain master's technique sheet, or hand it sample chapters of your own — what lands on the page has to be your sentence.

Style presets: Six technique sheets distilled from public-domain masters, swapped from beside the chat box.
A profile of your own: Upload your own sample chapters; it derives your style profile and writes to it from there on.
Style re-render: Draft already written, just wrong in the mouth? Re-polish the whole passage toward a target voice.
De-AI · edit memory: The phrasings you delete are remembered, and avoided in the next chapter.

The style preset panel open above the chat composer: Gothic Atmospheric, Ironic Social Observer, Lyrical & Luminous and Pulp Adventure listed in turn, each with a line of technique keywords beneath it — free indirect mockery, the unsaid as plot, image that becomes symbol; the composer's bottom row reads Editor, + File, Style, Learn my edits, Review chapter canon and Send — The ✎ Style layer beside the chat box · every preset shows its own technique

A writing workbench you can trust

Character relations, automatic context injection, canon and foreshadowing kept in order; edit by hand or by agent anytime, follow every chapter, and roll back any change in one click.

The worldbook workbench: characters, places and foreshadowing in 7 entries, retrieved on demand or always carried, with token totals and version history — The worldbook workbench · inject on demand · versioned · Chinese edition shown

The half-written book, moved in

Drop in one file — a book already hundreds of thousands of words long — and it splits into chapters, marks them as finished prior text, and builds the retrieval index. From then on continuations recall the earlier chapters for you; no more pasting the previous chapter across by hand. Deriving a starting worldbook, or a searchable summary per chapter, are switches: leave them off and they cost you nothing.

The manuscript import page with the split previewed and the import not yet started: found 9 chapters totalling 3,166 words at an average of 352 words per chapter, each chapter listed with its title and word count; of the four switches, mark as finished prior text and build RAG index are on, derive a starting worldbook and generate searchable chapter summaries are off; estimated usage reads under a minute, under 0.01 US dollars, 9 embedding calls and 0 AI calls, with a Start import button below — The split, previewed before you commit · 9 chapters, 3,166 words · 9 embedding calls · 0 AI calls

Editors on different models, at your service

Planning, drafting, polishing and research each get the right pen, dispatched by task; official APIs only, with usage and cost shown for every call.

Session cost breakdown: thirteen pipeline roles listed with their model and spend, totaling $1.31 — One real session's bill · by role, by model · Chinese edition shown

Bring your own agent

Loomnovel is also an MCP server. Write inside Claude Code or any agent you already use — manuscripts, worldbook, retrieval, the writer and the reviewing editor, de-AI polish: sixteen editorial tools, connected with a single message.

One message to connect: Sign in, issue an access token in Settings, and send the ready-made message to your agent — nothing to install, works with Claude Code, Cursor, or any MCP client.
The token is your identity: A token sees only your own books; usage shares the same daily allowance as the web app; revoke any time and it dies instantly.
One editorial room: The writer, the reviewer, and the de-AI pass your agent calls run through the same pipeline and the same story assets as the web app — both sides work on the same book.

Said plainly, up front

Your data, your sovereignty: We do not sell your writing, use it to train our own models, or expose it to other users. To run the features you choose, content passes only through the model and cloud processors disclosed on this page. Everything can be exported or deleted at any time, with processes built to GDPR and other applicable requirements.
No middleman markup: Model calls go through official APIs and audited cloud services only, never third-party relays. Token usage and cost are shown for every call, with no hidden multipliers. Models are applied by task, from budget-friendly to top-tier; prices reference each provider's public API pricing, and actual billing follows what this page discloses.
Chinese and English are two crafts: Each language gets its own writing assets and evaluation system: genre habits, syntax, aesthetics and ban lists built separately, not translated.

About Loomnovel

A vertical writing agent, built for one thing: long-form fiction in Chinese and English.

Product: Unlike continuation tools such as Sudowrite or NovelAI, Loomnovel is agentic in the real sense: in conversation, the author directs the model to call tools — research a subject, retrieve canon, revise the manuscript. Retrieval, cross-chapter memory, and deterministic context injection together keep very long serials consistent — exactly where general-purpose writing tools collapse past the first hundred thousand words.
Numbers: Production retrieval load-tested on a 346,000-character, 100-chapter corpus with 30/30 recall at 608ms p95; a four-million-character serial imported end-to-end with canon-accurate continuation; roughly $0.37 per chapter (measured competitor cost: $0.7–1.1); 145 published architecture decisions, 700+ automated tests, CI/CD with daily backups and restore drills. We only quote numbers we can reproduce.
Founder: Karl, Information Systems at the Technical University of Munich, has written serialized fiction with AI in depth since Gemini 2.5 Pro shipped, with a real serial record and reader following; he interned at an AI startup and independently built products during that time. Behind Loomnovel sits a million-character, verifiable writing workflow spanning every major frontier model. Our conviction: even as new frontier models keep arriving, the Gemini series remains unmatched for long-form fiction — each model has its own prose voice and instruction fidelity, and those differences can be measured. Chunhua (Karl) Wang on LinkedIn →
Compliance & trust: Fully EU-hosted, including observability; data is exportable and deletable at any time, reference texts are deleted within 24 hours, and multi-tenant isolation is live-tested; full Terms of Service online (GDPR / DSA / DMCA / EU AI Act framework, under legal review); zero tolerance for content involving minors; prompt-injection defenses hardened through two red-team rounds. We do not sell your content or train our own models on it.
Belief: Coding is not the only craft worth optimizing models for. Coding is scaffolding — we used it to integrate the full professional writing pipeline, so what ships is work an author can stand behind, not the AI slop found everywhere.
Where we stand: Market research, pricing research and business-viability analysis, a complete bilingual pipeline, interoperability with other agents, cross-platform portability, and data protection are in place, and the product is live with real users. What we need now is Google Cloud: to host our database and the services currently deployed on Railway, and to fully restore our deeply Gemini-integrated writing workflows.

Contact: [email protected]

Finish that story.

Beta seats open in batches. Write to us about the story you want to tell; your invite code ships the moment one frees up.

Start free trial Request an invite Have a code? Start now