high

A minimalistic, terminal-based interface for Large Language Models (LLMs) designed for low-latency feedback.

Motivation

Why Vibe Coding Fails - Ilya Sutskever

The major limitation of the current LLMs is the introduction of new bugs in the attempt to fix the target one(s). The only possible way out of it is to closely monitor what the LLM is doing, how does it thinking, what problems it might report (a missing source code header? or a specification file? or a tool-call declaration?) and in case if any problem appears, one could just do Ctrl + C and continue the conversation by typing high -C "clarifications...". Moreover, if one deals with a lot of code in the output of LLM the syntax highlighting simultaneously with streaming support are required. etc.

Most TUI LLM suites today are implemented extremely inefficiently. For example, charmbracelet/mods can delay final output by a minute or so. The Claude Code supports max 60 FPS due to the internal usage of React (and all that it implies). etc.

The high solves the problem of inefficiency with a Preview Mode. It streams raw tokens immediately as "ghost text" so one can read the output as it generates, then seamlessly swaps in the syntax-highlighted version the moment the line completes.

The high also solves the problem of bloat by basically having only three dependency: the highlight, the curl and the jansson (everything is available via the standard Debian packages)

For example,

time high -m Qwen3.5-397B-A17B-IQ2_KL -f -p --md "output a mandelbrot in colour into tui via awk /bash"

# The gif is obtained via:
# 
#ttyrec -e 'time high -m Qwen3.5-397B-A17B-IQ2_KL -f -p --md "output a mandelbrot in colour into tui via awk /bash"'
#ttygif ttyrecord
#gifsicle -O9 ./tty.gif -o tty.compressed.gif

This approach offers the control of real-time streaming with the polish of syntax highlighting, without the overhead of heavy TUI frameworks. The tool is basically a highly optimized clone of mods where all syntax highlighting work is delegated to highlight and all conversations are kept in json.

Installation

Build Dependencies

• Compiler: g++ or clang (C++20 support required).
• Libraries: libcurl, libjansson.

Runtime Dependencies

• highlight: Optional. If installed, high uses it for syntax highlighting. If missing, code blocks display as plain text.
• xclip: Optional. If installed, the ID of the selected conversation will be copied into the clipboard.
• fzf: Required for cb utility (codeblock extractor).
• bat or batcat: Optional. Used by cb for enhanced codeblock preview.

Setup

sudo apt install g++ build-essential libcurl4-openssl-dev libjansson-dev
sudo apt install highlight xclip bat fzf gawk
# 
make
sudo make install
# Use syntax highlighting with ghost text and markdown highlighting by default
high --format --preview --md --config-dump | tee -a ~/.bashrc
source ~/.bashrc
# Install extra utilities (pre, cb)
sudo make install-extras

Features

• Zero-latency Streaming: Preview mode displays tokens immediately with ghost text, swapping in highlighted versions on completion
• Smart Syntax Highlighting: Automatic language detection with fallback to markdown formatting (-f, --md)
• Conversation Management: Save, load, branch, and resume conversations with automatic timestamped titles
• Interactive Selection: TUI interface for browsing conversation history with vim-style keybindings (-l)
• Flexible Output Modes: Raw JSON (-r), plain text (-F), or formatted markdown
• Universal Piping: Reads from stdin and writes to stdout; integrates seamlessly into Unix pipelines
• Extra Utilities: pre for file context injection, cb for interactive codeblock extraction

Usage

Highlighting with streaming (--format --preview --md)

The highlighting (both for codeblocks and the rest of the output) is enabled by default. The --preview mode is optional, but essential especially for slower models (~20 tps decode).

1. Immediate Feedback: As the LLM emits tokens, they appear instantly in a dimmed color (ghost text).
2. Automatic Highlighting: When a code line completes, high passes it to the highlight binary and replaces the ghost text with the fully colored version.

The Flow

high reads from stdin and writes to stdout. It fits naturally into Unix pipelines.

curl -s https://github.com/ikawrakow/ik_llama.cpp/issues/629 | html2text | high -m Qwen-3.5 "what is the best hardware for the llm inference?"

Conversation management

high --models
> Kimi-K2.5-GGUF_IQ3_K
> Qwen3.5-397B-A17B-IQ2_KL
> Qwen3.5-397B-A17B-IQ4_KSS

# Ask Kimi-K2.5 to make a script
high -m Kimi-K2.5 "write a script for a full system backup into the .tar.gz archive"

# View the last conversation
high -S

# Continue the last conversation with a different LLM
high -C -m Qwen3.5 "dump it to a file and execute via bash /bash"
# Any time you can interrupt the conversation and add more context
# `Ctrl + C` -> `y`
cat ~/credentials.txt | high -C "use these to upload to the remote server"

# Branch conversation (save as new instead of continuing existing)
high -C -i -m Qwen3.5 "try a different approach"  # Creates new conversation branched from last

# System prompt manipulations etc.
high -m Qwen3.5 -R "you are professional system architect with 50y experience" "can you write a comprehensive chess engine with AI and a polished tui in c?"

History management

# Select a conversation from history (interactive TUI)
# (If xclip is installed also copies the id of the selected
# conversation into be clipboard)
high --list

# Alternatively, the non-interactive mode:
high --list | head -n 72 | tail 
> conv_1772224588366 (2026-02-28 10:27) [Qwen3.5-397B-A17B-IQ2_KL]
> conv_1772225893755 (2026-02-27 22:58) [Kimi-K2.5-GGUF_IQ3_K] [Interrupted]
> conv_1772224513458 (2026-02-27 22:35) [Qwen3.5-397B-A17B-IQ2_KL]
> conv_1772220106336 (2026-02-27 21:55) [Qwen3.5-397B-A17B-IQ4_KSS]
> conv_1772219943271 (2026-02-27 21:47) [Qwen3.5-397B-A17B-IQ2_KL]
> conv_1772221157687 (2026-02-27 21:39) [Kimi-K2.5-GGUF_IQ3_K]
> conv_1772219392149 (2026-02-27 21:09) [Kimi-K2.5-GGUF_IQ3_K] [Interrupted]
> conv_1772216271607 (2026-02-27 20:43) [Qwen3.5-397B-A17B-IQ4_KSS]
> conv_1772216654829 (2026-02-27 20:39) [Qwen3.5-397B-A17B-IQ2_KL]
> conv_1772217147864 (2026-02-27 20:32) [Kimi-K2.5-GGUF_IQ3_K] [Interrupted]

# Display the target conversation
high -s conv_1772224588366

# Plain text output for piping
high-plain -s conv_1771897941504 | less

# Extract user messages from raw JSON
high -s conv_1771897941504 --raw | jq -r '.messages[] | select(.role == "user") | .content' | grep --colour hello

Conversation Modes

Flag	Behavior
-c TITLE	Continue from saved conversation TITLE, append to same file
-C	Continue from last conversation, append to same file
-c TITLE -i	Branch from TITLE, save as NEW conversation
-C -i	Branch from last conversation, save as NEW conversation
(none)	Start fresh conversation with auto-generated title

Extra Utilities

After running sudo make install-extras, two additional utilities are available:

pre - File Context Provider

Injects file contents with proper markdown codeblocks into LLM prompts.

Features:

• Automatic language detection based on file extension
• Recursive directory scanning (maxdepth 2)
• Proper trailing newline handling
• Full path resolution

Usage:

# Single file
pre src/main.cpp | high "Review this code"

# Directory search
pre ./src "*.py" | high "Find all Python files and explain their purpose"

# Multiple files
pre src/*.*pp include/*.h | high "Analyze the architecture"

# Combine with conversation
pre docs/api.md | high -C "Update the documentation based on these changes"

Supported Extensions: sh, bash, zsh, py, js, ts, rb, go, rs, c, cc, cpp, h, hpp, java, cs, php, html, htm, css, json, xml, yaml, yml, md, sql

cb - Interactive Codeblock Extractor

Extracts codeblocks from high output (or any markdown source) with interactive selection.

Features:

• Parses markdown codeblocks (lang ... )
• Interactive TUI with fzf
• Preview with syntax highlighting (requires bat/batcat)
• Automatic clipboard copy (requires xclip)
• Shows block metadata (ID, line number, line count, language)

Usage:

# Extract codeblocks from conversation
high -S -F | cb

Output:

• Selected codeblock copied to clipboard (with xclip)
• Metadata displayed: block ID, language, line count
• Falls back to stdout if not in TTY mode

Dependencies:

• Required: fzf, awk
• Optional: bat/batcat (enhanced preview), xclip (clipboard copy)

Installing Extras

# Install to /usr/local/share/high and create aliases in /usr/local/bin
sudo make install-extras

# Verify installation
which pre
which cb

# Uninstall extras
sudo make uninstall-extras

Adding to PATH

Add to ~/.bashrc or ~/.zshrc:

export PATH="/usr/local/bin:$PATH"

Environment Variables

Variable	Description	Default
LLM_API_BASE	API endpoint URL	http://localhost:8042/v1
LLM_CONFIG_DIR	Configuration directory	~/.config/high
LLM_SYSTEM_ROLE	Default system prompt	You are a helpful assistant.
LLM_MAX_TOKENS	Maximum tokens for response	0 (unlimited)
LLM_TEMPERATURE	Sampling temperature	1.0
LLM_TIMEOUT	Request timeout in seconds	36000
LLM_FORMAT_MARKDOWN	Enable markdown formatting	1
LLM_VERBOSE	Verbose output	0
HIGHLIGHT_THEME	Syntax highlighting theme	xoria256
HIGH_PREVIEW_CODE	Enable code preview mode	0
HIGH_MARKDOWN_SYNTAX	Enable markdown highlighting	1
LLM_TAB_WIDTH	Tab expansion width	8
HTTP_PROXY	HTTP proxy URL	(none)

Examples

# Basic usage
high "What is the capital of France?"

# With specific model
high -m llama-3 "Explain quantum computing"

# Continue last conversation
high -C "Add more details about..."

# Branch from last conversation (create new)
high -C -i "Try a different approach"

# List conversations
high -l

# Show conversation with full highlighting and a ghost-text
high -s myconv -f -p --md

# Pipe from curl
curl -s https://example.com/api | high "Summarize this"

# Export configuration
high --config-dump >> ~/.bashrc

# Debug log
high -S -f -p --md --debug --log-file high_debug.log

# Use with pre utility
pre src/*.cpp | high "Review this code for security issues"

# Extract codeblocks from conversation
high-plain -S | cb

# List and apply the specific highlight theme
highlight --list-scripts=themes | grep -F ': vim'
high --theme bluegreen -S -f -p --md

# "Feature: chat start-new should return or display new chat ID"
# https://github.com/charmbracelet/mods/issues/675
# 
NEW_ID=$(high -m Qwen3.5 "hello" 2>/dev/stdout | grep -Eo 'conv_[0-9]+' | tail -n1)
pre file.txt | high -c "$NEW_ID" --isolate "Message"

Files

Path	Description
~/.config/high/conversations/	Directory containing saved conversations (JSON format)
/usr/local/share/high/	Extra utilities (pre, cb)
/usr/local/bin/pre	Alias script for pre utility
/usr/local/bin/cb	Alias script for cb utility

Troubleshooting

No syntax highlighting

• Ensure highlight is installed: sudo apt install highlight
• Check theme availability: highlight --list-scripts=themes
• Use --theme to specify different theme

Conversation not saving

• Check permissions: ls -la ~/.config/high/conversations/
• Verify disk space
• Check for JSON parse errors in logs

Extras not working

• Verify installation: sudo make install-extras
• Check PATH: echo $PATH
• Verify dependencies: which fzf, which batcat

See Also

• curl(1) - HTTP client
• highlight(1) - Syntax highlighting
• jq(1) - JSON processor
• fzf(1) - Fuzzy finder
• bat(1) - Cat with syntax highlighting

Bugs

Report bugs at: https://github.com/magikRUKKOLA/high/issues

Author

magikRUKKOLA magikRUKKOLA@proton.me

License

GPL-3.0