Two local minds, one collaborative story.
Storymaker lets two local Ollama models take turns writing a story. You provide the opening line, pick a model for each author, and watch the prose stream in live as the two models hand the narrative back and forth. The finished story can be narrated aloud with a local Kokoro text-to-speech voice.
Storymaker comes in two flavors that share the same engine and UI:
- Copilot CLI extension — opens a native desktop window from inside GitHub Copilot CLI. Lives in
.github/extensions/storymaker/. See Installation below. - Standalone web app — a self-contained local server that runs in your browser with plain
node server.mjs, no Copilot CLI needed. Lives instandalone/. See standalone/README.md.
Both produce identical results; pick whichever fits your setup. The rest of this README describes the extension.
- 🤝 Two-author collaboration — Author A writes a configurable number of sentences, Author B continues, and they alternate for as many turns as you choose.
- 🏁 Optional conclusion — on by default, the opening author (Author A) wraps up the story with a closing passage of configurable length (1×–4× the per-turn sentence count, default 2×).
- 🧠 Full shared context — every turn, each author receives the entire story so far, so the narrative stays coherent as it grows.
- 🎛️ Pick any installed model per author — both dropdowns are populated from your local Ollama models; mix and match (e.g.
llama3.2+qwen2.5). - ⚡ Live token streaming — text appears word-by-word with a blinking caret, color-coded by author.
- ⏹️ Stop anytime — cancel mid-generation and keep the partial story.
- 📋 Copy & save — copy the finished story to the clipboard, or save it to a file that begins with a YAML front matter block recording the models and settings used.
- ⌨️ Editing shortcuts — select-all, copy, cut, paste, and undo work in the text fields.
- 🎨 Audible-inspired dark theme — near-black canvas, signature orange accents, serif story type.
┌──────────────────────────┐ ┌──────────────────────────┐
│ Webview (the UI) │ RPC │ Extension (main.mjs) │
│ pick models, see story │ ◄────► │ orchestrates the loop │
└──────────────────────────┘ └────────────┬─────────────┘
│ HTTP
▼
┌──────────────────────┐
│ Ollama (localhost) │
│ /api/tags /api/chat │
└──────────────────────┘
The extension (Node) owns all model orchestration. For each turn it sends the full story so far to the chosen model via Ollama's streaming /api/chat, appends the result, and pushes live tokens to the page. The page never talks to Ollama directly — it calls back into the extension through the webview bridge (window.copilot.*) and receives updates via window.sm.*.
Reasoning is disabled (think: false) so the models write prose directly instead of spending their token budget thinking — with an automatic fallback for models that don't support that flag.
- GitHub Copilot CLI
- Ollama running locally (default
http://localhost:11434) - At least one pulled model — any model shown by
ollama listworks, for example:ollama pull llama3.2 ollama pull qwen2.5
Privacy: By default Storymaker talks to a local Ollama instance, so your starter text and the generated story stay on your machine. If you point
OLLAMA_HOSTat a remote or LAN server, that text is sent there instead.
Storymaker is a project-scoped Copilot CLI extension. Clone this repo and open Copilot CLI from inside it:
git clone https://github.com/mmacy/storymaker.git
cd storymaker
copilotThe extension lives in .github/extensions/storymaker/ and is auto-discovered. On first load Copilot installs its npm dependencies automatically. If the extension doesn't appear, run /reload (or restart the CLI).
From within Copilot CLI:
/storymaker
This opens the Storymaker window. Then:
- Enter starter text — the opening of your story.
- Choose a model for Author A and Author B.
- Set sentences per turn (1–10) and number of turns (1–25).
- Optionally let Author A write a conclusion (on by default) and pick its length (1×–4× the per-turn sentence count).
- Click Weave story and watch the two authors collaborate.
- Copy or Save the finished story (saved files land in the repo root by default).
The agent can also open and drive the window via the
storymaker_show,storymaker_eval, andstorymaker_closetools.
| Setting | Where | Default |
|---|---|---|
| Ollama endpoint | OLLAMA_HOST env var |
http://localhost:11434 |
| Sentences per turn | UI | 2 (range 1–10) |
| Number of turns | UI | 3 (range 1–25) |
| Conclusion | UI | On |
| Conclusion length | UI | 2× sentences-per-turn (range 1×–4×) |
| Window size | main.mjs (width/height) |
1588 × 1280 (physical px) |
Saved stories begin with a YAML front matter block capturing how the story was generated, followed by a blank line and the story text:
---
title: "The venerable Mazpar awoke."
generator: Storymaker
created: "2026-06-04T21:12:11.611Z"
author_a_model: "qwen3.6:35b"
author_b_model: "gemma4:latest"
sentences_per_turn: 2
turns: 5
conclusion: true
conclusion_length: "2x"
conclusion_sentences: 4
ollama_host: "http://localhost:11434"
---Copying to the clipboard yields the plain story without front matter.
.github/extensions/storymaker/ # the Copilot CLI extension
├── extension.mjs # bootstrapper (installs deps, imports main.mjs)
├── main.mjs # Ollama client + generation loop + callbacks
├── package.json
├── lib/ # reusable webview library (do not edit)
│ ├── copilot-webview.js
│ └── webview-child.mjs
└── content/ # the UI served to the window
├── index.html
├── style.css
└── main.js
standalone/ # the standalone browser app
├── server.mjs # HTTP + SSE server with the same engine
├── package.json
├── README.md
└── content/ # verbatim copy of the extension's content/
See AGENTS.md for contributor and AI-agent guidance.
Released under Creative Commons CC0 1.0 Universal — a public-domain dedication. You can copy, modify, and distribute this work, even for commercial purposes, without asking permission.
