1667 0.9.8 Release: Terminal UI for Drafting Fiction with Language Models

1667 0.9.8 lets you write fiction in the terminal while keeping every AI‑generated draft in a navigable tree

Takeaway: 1667 version 0.9.8 provides a self‑contained, open‑source terminal UI that prompts language models for the next paragraph, stores every generated take side‑by‑side in a tree, and does so without any telemetry, accounts, or cloud sync.

Core workflow: generate, compare, and shape the story line

Prompt‑driven generation: Press Enter to open the composer, type a direction (e.g., "let the unlit lantern answer"), and the model returns a paragraph. The program then pauses, waiting for the writer’s next command.
Take preservation: Each generated paragraph is saved as a take. Using ← and → flips between sibling takes; the original take remains untouched. Only the d command deletes a take and its descendants.
Tree‑based navigation: The story is represented as a tree where each node is a paragraph and each branch is an alternative take. Press M to open a map view that visualizes the entire tree, showing sibling circles, word counts, and the current path.
File‑system integration: Stories live in a plain‑text folder chosen by the user. Export writes the selected path to Markdown in the same directory; no server or account is required.

Privacy‑first design: no analytics, no telemetry, no hidden servers

"Every request the program makes goes to the endpoint you configured. Nothing comes to us, because there is no server of ours to send it to." – project documentation

Zero data collection: The source contains no analytics library, telemetry hooks, crash‑reporter endpoints, or install IDs.
Optional update check: The only outbound request is an optional version‑check that reads a version number; it never downloads or installs anything.
Local key storage: API keys are stored in a private file outside the story folder and are never transmitted to the 1667 project maintainers.
Apache‑2.0 license: All code is openly available on GitHub, allowing anyone to audit the privacy claims.

Supported platforms and installation

Platforms: macOS (arm64/x64), Linux (glibc 2.17+ arm64/x64), Windows (x64 PowerShell installer).

Installation command:

curl -fsSL https://1667.ai/install.sh | sh

Verified installer: Users can verify the release with GitHub CLI attestation before running the script, ensuring the binary matches tag v0.9.8.

Model flexibility and configuration

Provider‑agnostic: 1667 does not bundle a model. It can point to any OpenAI‑compatible, Anthropic‑compatible, or self‑hosted endpoint (e.g., Ollama, LM Studio, llama.cpp, KoboldCpp).
Editable settings: Press , to open the generation settings panel where you can adjust temperature, max tokens, context window (default 32,768 tokens), cache policy, and more.
Dry‑run mode: The UI includes a dry‑run provider for testing without contacting an external API.

Keyboard‑centric UI: all actions in a few keys

Category	Key(s)	Action
Move	↑ ↓	Navigate previous/next row
	← →	Flip between sibling takes
	G / Shift+G	Jump to first/last part
Write	Enter / I	Open composer for a new prompt
	Space	Continue current paragraph
	R / Shift+R	Retake with same or edited prompt
	W	Write a take manually
Open	M	Open map view
	F	Open facts panel
	O	Switch story or open library
Shape	D	Delete current take and its subtree
	T	Tag the current line
	C / Shift+C	Create or end a chapter
Map	A	Show all takes or sketches
	L	Follow a tree line
	S	Cycle mass sorts

The full reference is available with ?.

Community reaction on Hacker News

Positive notes: Users praised the beautiful TUI design and the novel tree‑based approach to preserving every AI‑generated draft. One commenter called it "really cool" and expressed interest in building similar tools.
Criticism of AI‑driven writing: Several commenters argued that relying on language models defeats the purpose of creative writing, labeling the tool as "depressing" or "spectacularly wrong‑headed".
Feature wishes: Some users suggested adding AI‑assisted brainstorming, feedback, or a character bible rather than pure paragraph generation. Others wanted LaTeX/EPUB export and richer visual feedback.
Technical concerns: A question was raised about the context window size (32 k tokens) and whether later chapters are sent to the model; the documentation states the user controls the endpoint and context.
Support for the approach: A handful of commenters defended the tool, emphasizing that the software itself is neutral and that open‑source AI‑assisted creativity deserves experimentation.

Frequently asked questions (from the project FAQ)

Does 1667 write the whole book for me? No. It only generates a paragraph when you explicitly request it; you decide which take to keep.
What happens to unused takes? They remain in the tree and are visible in the map view; only the d command removes them.
Is any personal data collected? No analytics, telemetry, or install IDs are present. The only external call is an optional version check.
Can I run it on Windows? Yes, via the PowerShell installer; the same install.sh command works on macOS and Linux.
What does it cost? The software is free under Apache‑2.0. You only pay for the model provider you connect to, if any.

Why the name “1667”?

The name references the goal of writing 1,667 words per day for 30 days, which yields a 50 k‑word first draft—a common target for novelists. The tool itself does not track whether you meet that goal.

Bottom line

1667 0.9.8 delivers a minimalist, privacy‑preserving terminal UI that treats every AI‑generated paragraph as a first‑class, versioned artifact. By storing the entire draft tree locally and exposing a rich set of keyboard shortcuts, it gives writers fine‑grained control over AI assistance while avoiding the pitfalls of cloud‑based, telemetry‑heavy alternatives. Whether you view AI‑generated prose as a creative partner or a shortcut, 1667 provides a transparent, open‑source platform for experimenting with language‑model‑driven fiction.