1667, a terminal UI for writing fiction with language models, was posted to Hacker News as a Show HN item on August 17, 2026, with a linked homepage at 1667.ai and source code on GitHub. The page describes 1667 as a tool that asks a configured language model for the next paragraph, then keeps every generated take side by side in a navigable tree. The Hacker News thread listed a score of 30 and 80 comments.

1667 0.9.8 installs through curl on macOS and Linux, PowerShell on Windows x64, npm as @1667-ai/cli@latest, or source from github.com/1667-ai/1667. The managed installer supports macOS arm64/x64 and glibc 2.17+ Linux arm64/x64; the npm path needs Node 22, and the source path needs Bun 1.3.14 or later plus Node 22. The page also gives optional installer verification commands using GitHub CLI, GitHub release download, and attestation verification for v0.9.8.

The writing workflow centers on a manuscript tree. Pressing Enter opens a composer, the writer types a direction, and 1667 generates the next story part before stopping. Up and Down move through story parts, Left and Right compare sibling takes, M opens a path map, and ? opens the keyboard reference. The demo story, “The lantern keeper,” shows chapter three, parts 11 through 13, with part 12 on take 3 of 5; a submitted direction, “let the unlit lantern answer,” produces the next part: “The lantern flame bent toward the compass, though no door had opened.”

The map view has path, tree, and mass modes. In the path map demo for “The lantern keeper,” depth 1 through 13 contains 23 parts across 4 story lines, with sibling takes beside rows 3, 5, 8, 11, and 12; part 12 is selected at take 3 of 5, and part 13 carries the canon-storm tag. In the tree map, the story contains 4 lines, 23 parts, and 5 forks. In the mass map, the story contains 477 words across 4 lines, with horizontal bars comparing canon-storm, alt-quiet-inn, burned, draft-ledger, seven sketches, and cold single-paragraph branches by word count.

The project says it has no analytics, no telemetry, no crash reporter, no install id, and no account system. It says stories stay in a chosen local directory, exports write Markdown next to them, and prose goes only to the provider or local endpoint selected by the user. The page says every request goes to the configured endpoint, with one exception: an optional update check that ships off, reads a version number if enabled, downloads nothing, and replaces nothing.

1667 includes no model and does not rent access to one. Its generation settings demo lists Provider: dry-run, Temperature: 0.7, Maximum tokens: 2,048, Context window: 32,768, Cache policy: off, and editable routes for defaults and prose. The models section lists OpenAI, Anthropic, OpenRouter, Ollama, LM Studio, llama.cpp, KoboldCpp, and Custom endpoint as destinations. The FAQ says 1,667 words a day for thirty days equals fifty thousand words, and the name refers to that first-draft pace, though the program does not track whether the writer hits it.

1667, a terminal UI for writing fiction with language models, was posted to Hacker News as a Show HN item on August 17, 2026, with a linked homepage at 1667.ai and source code on GitHub. The page describes 1667 as a tool that asks a configured language model for the next paragraph, then keeps every generated take side by side in a navigable tree. The Hacker News thread listed a score of 30 and 80 comments.

1667 0.9.8 installs through curl on macOS and Linux, PowerShell on Windows x64, npm as @1667-ai/cli@latest, or source from github.com/1667-ai/1667. The managed installer supports macOS arm64/x64 and glibc 2.17+ Linux arm64/x64; the npm path needs Node 22, and the source path needs Bun 1.3.14 or later plus Node 22. The page also gives optional installer verification commands using GitHub CLI, GitHub release download, and attestation verification for v0.9.8.

The writing workflow centers on a manuscript tree. Pressing Enter opens a composer, the writer types a direction, and 1667 generates the next story part before stopping. Up and Down move through story parts, Left and Right compare sibling takes, M opens a path map, and ? opens the keyboard reference. The demo story, “The lantern keeper,” shows chapter three, parts 11 through 13, with part 12 on take 3 of 5; a submitted direction, “let the unlit lantern answer,” produces the next part: “The lantern flame bent toward the compass, though no door had opened.”

The map view has path, tree, and mass modes. In the path map demo for “The lantern keeper,” depth 1 through 13 contains 23 parts across 4 story lines, with sibling takes beside rows 3, 5, 8, 11, and 12; part 12 is selected at take 3 of 5, and part 13 carries the canon-storm tag. In the tree map, the story contains 4 lines, 23 parts, and 5 forks. In the mass map, the story contains 477 words across 4 lines, with horizontal bars comparing canon-storm, alt-quiet-inn, burned, draft-ledger, seven sketches, and cold single-paragraph branches by word count.

The project says it has no analytics, no telemetry, no crash reporter, no install id, and no account system. It says stories stay in a chosen local directory, exports write Markdown next to them, and prose goes only to the provider or local endpoint selected by the user. The page says every request goes to the configured endpoint, with one exception: an optional update check that ships off, reads a version number if enabled, downloads nothing, and replaces nothing.

1667 includes no model and does not rent access to one. Its generation settings demo lists Provider: dry-run, Temperature: 0.7, Maximum tokens: 2,048, Context window: 32,768, Cache policy: off, and editable routes for defaults and prose. The models section lists OpenAI, Anthropic, OpenRouter, Ollama, LM Studio, llama.cpp, KoboldCpp, and Custom endpoint as destinations. The FAQ says 1,667 words a day for thirty days equals fifty thousand words, and the name refers to that first-draft pace, though the program does not track whether the writer hits it.