bito-lint
bito = building in the open.
AI coding agents generate documentation as they work — ADRs, design docs, changelogs, handoff notes. The quality varies between sessions. Sometimes you get crisp, well-structured prose. Sometimes you get bloated walls of text that no one wants to review.
bito-lint catches the problems before you commit. It runs 18 deterministic writing checks — readability scoring, token budgets, section completeness, grammar, dialect enforcement, and style analysis. No LLM, no API calls, no network. Same input, same result, every time.
The goal: agent-generated documents that are clean enough to ship.
$ bito-lint analyze docs/architecture.md
docs/architecture.md
Readability: Grade 12.4, 24 sentences, 390 words
Grammar: 8 issues, 3 passive (12.5%)
Sticky: Glue index 21.5%, 2 sticky sentences
Pacing: Fast 62% / Medium 29% / Slow 8%
Length: Avg 15.8 words, variety 10.0/10
Transitions: 0% of sentences, 0 unique
Overused: "template" (1.3%), "skill" (1.1%), "design" (1.1%)
Diction: 2 vague words
Style: Score 92/100, 2 adverbs, 0 hidden verbs
What it checks
analyze runs 18 checks in one pass:
| Category | What it catches |
|---|---|
| Readability | Flesch-Kincaid grade level — flag documents that demand too much of the reader |
| Grammar | Passive voice, double negatives, subject-verb disagreement, missing punctuation |
| Sticky sentences | High "glue word" density — sentences stuffed with is, the, of, in |
| Pacing | Monotonous sentence rhythm — all short punches or all long slogs |
| Sentence variety | Length distribution — a score of 1/10 means every sentence is the same length |
| Transitions | Percentage of sentences using connective phrases — low means choppy reading |
| Overused words | Repeated non-trivial words that make the text feel circular |
| Repeated phrases | Bigrams and trigrams that recur too often |
| Echoes | Same word appearing in adjacent sentences (unintentional repetition) |
| Complex paragraphs | Paragraphs with too many ideas competing for attention |
| Conjunction starts | Sentences opening with But, And, So — fine in moderation, a tic in excess |
| Cliches | "At the end of the day," "move the needle," "low-hanging fruit" |
| Diction | Vague words (things, stuff, very) that add length without meaning |
| Sensory language | Percentage of concrete, sensory words — useful for judging descriptive writing |
| Consistency | Mixed US/UK spelling (color and colour in the same document) |
| Dialect enforcement | Flag spellings that violate your project's chosen dialect (en-us, en-gb, en-ca, en-au) |
| Acronyms | Tracks acronym usage for consistency |
| Style score | Combined metric: adverb density, hidden verbs (nominalizations), overall polish |
Every check is deterministic. No API calls, no LLM, no network. The same input produces the same output every time.
Focused checks run individually when you need a specific gate:
# Does this handoff fit in 2,000 tokens?
)
# Is this user guide accessible to a general audience?
)
# Does this ADR have all the sections it needs?
)
# How's the grammar?
)
Installation
Homebrew (macOS and Linux)
From source
Pre-built binaries
Download from the releases page. Binaries are available for macOS (Apple Silicon and Intel), Linux (x86_64 and ARM64), and Windows.
Usage
Full analysis
Add --json for machine-readable output. Add --dialect en-gb to enforce British spelling. Add --checks readability,consistency to run only specific checks.
Quality gates
Quality gates are pass/fail checks designed for CI, pre-commit hooks, and automation:
# Token counting with budget enforcement
# Readability with grade ceiling
# Section completeness against a template
# Grammar and passive voice analysis
Built-in completeness templates: adr, handoff, design-doc. Define your own in a bito-lint config file.
Every command exits non-zero on failure, writes structured JSON with --json, and works in pipes.
Dialect enforcement
Set a project dialect and bito-lint flags wrong-dialect spellings alongside mixed-spelling inconsistencies:
# Via flag
# Via environment variable
# Via config file (.bito-lint.toml)
Supported dialects: en-us, en-gb, en-ca (Canadian hybrid: US -ize/-ise, UK for the rest), en-au.
MCP server
bito-lint includes a built-in MCP server, so AI coding assistants can call quality gates directly during writing sessions:
This exposes six tools: analyze_writing, count_tokens, check_readability, check_completeness, check_grammar, and get_info. Total schema cost: ~1,283 tokens. See docs/mcp-development.md for context budget details.
Configuration
Drop a config file in your project and it takes effect automatically:
.bito-lint.toml(or.yaml,.json) in the current directory or any parentbito-lint.toml(without dot prefix) in the current directory or any parent~/.config/bito-lint/config.toml(user-wide defaults)
Closer files win. All formats (TOML, YAML, JSON) work interchangeably.
# .bito-lint.toml
= "en-us"
= 2000
= 12.0
= "warn"
Shell completions
Included in Homebrew installs and release archives. For manual setup:
# Bash
# Zsh
# Fish
Development
crates/
├── bito-lint/ # CLI binary
└── bito-lint-core/ # Core library
Prerequisites: Rust 1.88.0+, just, cargo-nextest.
Commit messages follow Conventional Commits. The project enforces safe Rust (#![deny(unsafe_code)]), clippy nursery lints, and cargo deny for dependency auditing.
License
MIT (LICENSE-MIT)