Inkhaven (blackInkhaven)
Inkhaven is a standalone terminal application for writing books and long-form technical documentation. It pairs a full-screen Typst editor with a local semantic index, an AI writing assistant, versioned snapshots, and a backup pipeline — so the entire writing workflow lives inside one binary, without leaving the terminal.
Your manuscript is organised as a hierarchy of .typ files
(Book → Chapter → Subchapter → Paragraph), with first-class
image (.png / .jpg / …), HJSON data (.hjson), and
Bund script (.bund) leaves alongside paragraphs. Inkhaven
stores metadata in a local DuckDB database, indexes every text
node for full-text and semantic search, keeps versioned
snapshots, embeds the Bund
scripting language for hooks + custom rules, and streams answers
from your chosen LLM provider — six are bundled (Gemini,
Claude, OpenAI, DeepSeek, Grok, Ollama) and any
model genai routes is
one HJSON line away.

Latest release · 1.2.23 — Stability & fixes
Read the full notes: Documentation/RELEASE_NOTES/1.2.23.md
The last 1.2 housekeeping pass before the 1.3 editorial pass cycle — no new headline feature, just a calmer, harder base: a crash-behaviour fix, a whole-tree panic-surface re-audit, SQL hardening, and the deferred find & replace polish.
Fixes & hardening
- Broken pipe is no longer a crash.
inkhaven … | headused to write a spurious crash-report file (Rust ignores SIGPIPE, so the failed write panicked); the panic hook now exits cleanly on a broken stdout pipe, while leaving the long-running TUI's pipe handling untouched. - Panic-surface re-audit across the whole tree fixed three reachable
UTF-8 byte-boundary slice panics — theme-colour parsing, fact-check
cursor jump, and the doctor naming scan, all triggerable on
multilingual or malformed input — and routed four
storewrites through the atomic-write path. - SQL hardening. The JSON / blob and progress stores now bind
parameters instead of string-interpolating; both
sql_escapecopies are gone, pinned by an adversarial round-trip test.
Find & replace polish
Inside the project-replace review modal (1.2.22), w / i / x now
toggle whole-word / ignore-case / regex in place — the same matching modes
the CLI flags reach — starting from the safe whole-word default.
Test stats
Tests 1186 → 1190, zero new Rust dependencies.
Every prior release lives under
Documentation/RELEASE_NOTES/.
Why Inkhaven
- Terminal-first. Inkhaven runs over SSH, in tmux, on a tiling WM — no browser, no Electron. The TUI uses ratatui and tui-textarea.
- Your manuscript is plain files. A paragraph lives in a
.typfile on disk; the metadata database tracks hierarchy and search but the prose is text — you can read it, diff it, version-control it, and edit it with another tool any time. - Semantic search out of the box. Embeddings via fastembed and HNSW are computed locally. Search for "the moment the lighthouse fails" and find the paragraph even if it never uses those exact words.
- AI is a co-author you steer. Inferences stream live; you control the
scope (selection / paragraph / subchapter / chapter / book), the
mode (Local-only RAG vs. Full general knowledge), and the
destination (replace, insert, top, bottom, copy, grammar-apply).
Inkhaven does NOT provide inherent privacy when external providers
(Gemini / Claude / OpenAI / DeepSeek / Grok) are used — prompts
travel to their servers under their terms. For privacy, set
llm.default_provider: "ollama"and run a local model; every other inkhaven subsystem (RAG embedding, semantic search, snapshot diff) is already on-device. - Multilingual. Snowball stemmers and multilingual embeddings make Russian, German, French, Spanish, Italian and others first-class. The shipped defaults cover English and Russian.
- Help, characters, places, artefacts, scripts — built in. Nine
system books are seeded on every project:
Notes,Research,Prompts,Places,Characters,Artefacts,Typst,Scripts,Help. Mentions of names from the lexicon books light up in the editor (cyan / amber / peach / underline).Ctrl+B P/C/Y/Gquery each via RAG.F1answers questions about Inkhaven itself by RAG overHelp.Scripts(added in 1.2) holds.bundsource files auto-loaded into the embedded Bund scripting VM at project open — seeDocumentation/Bund/. - First-class images. Drop PNG / JPG / WebP / SVG into the tree;
Book assembly emits the right
wrap_image_*calls and ships the bytes into the typst tree.Ctrl+B Pinside#image("…")opens a sibling picker. Enter on an Image row pops a ratatui-image preview (kitty / sixel / iterm2 / half-block). - From buffer to PDF in two chords.
Ctrl+B Aassembles your tree into a typst-compilable directory;Ctrl+B Bcompiles it;Ctrl+B Obuilds and copies the PDF into your shell's cwd as<book>-YYYYDDMM-HHMM.pdf. Compile failures route the captured stderr into a fresh AI chat with a typst-aware system prompt.
Features at a glance
Editor
- Typst syntax highlighting via tree-sitter-typst.
- Regex find / replace with same-line current-match highlighting.
- Split-edit with versioned snapshots — see two versions of a paragraph side by side, accept either.
- Word-aware navigation and deletion shortcuts.
- Vertical block selection (Alt+arrows) with rectangular copy.
- System-clipboard cut / copy / paste, plus per-doc undo / redo.
- Live "changes since last save" bolding; grammar-correction highlights
what changed after a
gapply.
Tree
- Multi-level folding (
←/→/Z/X). - Per-kind row colours (book / chapter / subchapter / paragraph / image)
- open-paragraph marker.
- Plain-letter shortcuts for add (
B/C/V/A/S/+/P), delete (D/-), reorder (U/J). - Document status badge column — one character per row colour-
coded to the workflow stage (
n/1/2/3/F/R). - Mouse: click to focus + select; scroll wheel scrolls.
AI pane
- Streaming markdown rendering — bold / italic / headings / code / lists.
- Six scope modes (cycled by
F9): None, Selection, Paragraph, Subchapter, Chapter, Book — each prepends the matching content to the next prompt. - Two inference modes (
F10): Local (use only supplied context) and Full (augment with general knowledge). Help inferences are pinned to Local automatically. - Persistent chat history with one-key clear (
Ctrl+B C). - Full-screen AI layout (
Ctrl+B K) — AI pane + scrollable chat history + AI prompt; persisted to.inkhaven-chat.jsonbetween sessions;Ctrl+Fsearches;Ctrl+Centers a turn-selection mode. - Lexicon RAG —
Ctrl+B P/C/G/Yin the editor sweep the selection throughPlaces/Characters/Notes/Artefactsand prepend the lookup to the next AI prompt. - F1 Help-manual floating query → grounded answer over the Help book.
inkhaven import-typst-helpseeds Help with a curated typst reference. - F7 Grammar check with deterministic correction extraction (
greplaces the buffer with just the corrected text, preserving Typst markup).
Storage and backup
- DuckDB metadata + DuckDB blobs + HNSW semantic vectors. No inverted full-text index — search is semantic-only via embedded vectors.
- Snapshots:
F5captures the buffer;F6opens the snapshot history picker. inkhaven backup --out <dir>zips the entire project.inkhaven restore <archive> --to <dir>puts it back.- Auto-backup on TUI exit when the last backup is older than
backup.max_age(humantime:7d,12h,30m, …) — splash screen with a progress bar. - Session persistence: cursor position, focus, tree-scroll, open paragraph all survive restarts. Per-paragraph cursor memory: switch around and every paragraph remembers where you were.
CLI tools
init— set up a fresh project (interactive confirmation if the directory exists).add/delete/mv/list— manage the hierarchy from a script.search "phrase"— semantic search from the shell.reindex— re-walk.typfiles into the database.export typst/export pdf— produce a single Typst manuscript or a built PDF.import-help --documents-directory <dir>— populate the Help book from a directory of markdown / text / typst files (wipes Help first).backup/restore— see above.ai "prompt"— one-shot inference from the shell (no TUI).
Configuration
A single inkhaven.hjson in each project root drives every knob:
embedding model, LLM providers, autosave cadence, sync interval, hierarchy
depth, language, snowball stemmers, the full visual theme (per-pane
backgrounds and foregrounds, all syntax colours, lexicon highlight
colours), key bindings, and backup policy.
Install
Inkhaven ships as a single static binary per platform. Three install paths:
1. cargo binstall (no compile)
If you already have cargo-binstall:
cargo-binstall reads [package.metadata.binstall] from Cargo.toml,
picks the right asset off GitHub Releases, and drops the binary into
~/.cargo/bin. Works on Linux (x86_64), macOS (Intel + Apple Silicon),
and Windows (x86_64).
2. GitHub Releases (direct download)
Grab the tarball for your platform from
Releases, unpack,
and put inkhaven somewhere on your PATH. Builds are produced by
the release.yml workflow on every
tag push.
3. cargo install inkhaven (compile from crates.io)
Inkhaven is published on crates.io — every release tag pushes a
new version (latest: 1.2.19). The first build takes ~10 minutes on
a modern laptop because of DuckDB + fastembed + ONNX-runtime
compilation; cargo binstall above is the fast path.
4. cargo install --git (compile from a specific tag)
Useful when you want a specific tag, a pre-release branch, or a local fork.
Quick start
# Build (if installing from source)
# Initialise a project (asks for confirmation if the directory exists)
# Build the hierarchy from the CLI…
# …or skip the CLI and add everything from the TUI
# Inside the TUI: B (book), C/V (chapter), A/S (subchapter), +/P (paragraph)
Use cases
- Long-form fiction. Hierarchy fits novels naturally (Book → Part → Chapter → Scene). Places / Characters / Research system books keep worldbuilding next to prose.
- Technical documentation. Each chapter is a
.typfile; the tree doubles as a table of contents. Semantic search makes "where did I document the retry policy?" a one-keystroke question. - Translation work. Multilingual embeddings + per-language Snowball stemmers let you keep source and target in two parallel books.
- Research notebooks. Snapshots track how a draft evolved; the AI pane can summarise a chapter when you come back after a week.
- Help and onboarding writing. Ship docs as a directory and let Inkhaven build a Help book your readers can query through F1.
Documentation
The full docs live under Documentation/.
Start here:
Documentation/README.md— entry point and table of contents.Documentation/FIRST_STEPS.md— compile, install, initialise.Documentation/Tutorials/— narrative walk-throughs, each focused on one workflow.
Reference:
Documentation/KEYBINDING.md— every keystroke the TUI recognises, organised by pane and overlay.Documentation/CONFIGURATION.md— the full HJSON reference.Documentation/MAINTENANCE.md— backup, restore, reindex, logs.Documentation/PROMPTS.md— the prompt library and the Prompts system book.Documentation/LOCATIONS.md— managing Places.Documentation/CHARACTERS.md— managing Characters.
Built with
- duckdb — metadata + blob persistence
- vecstore — HNSW semantic vector store
- fastembed — embedding model (search is semantic-only; no inverted full- text index)
- ratatui, tui-textarea
- tree-sitter + tree-sitter-typst
- genai — provider-neutral LLM streaming
- pulldown-cmark,
rust-stemmers,
zip,
humantime, and many others —
see
Cargo.toml.
Licence
Apache 2.0 — see LICENSE.
Security Warning and Disclaimer
Inkhaven is provided "AS IS" with no warranty. The author cannot be held liable for personal, business, or financial damage arising from its use. Use is voluntary and at your own risk.
Before opening a project file you did not author — and before
relying on Inkhaven for work you cannot afford to lose —
please read Documentation/SECURITY_WARNING.md.
It enumerates the security issues catalogued in the 1.2.15
audit (both fixed and pending), the design properties that
are inherent rather than bugs, the unknown-risk classes that
no audit can fully eliminate, and the limitation-of-liability
terms under which Inkhaven is distributed.