Chatnik

Raku package that provides Command Line Interface (CLI) scripts for conversing with persistent Large Language Model (LLM) personas.

"Chatnik" uses files of the host Operating System (OS) to maintain persistent interaction with multiple LLM chat objects.

"Chatnik" can be seen as a package that "moves" the LLM-chat objects interaction system of the Raku package "Jupyter::Chatbook", [AAp3], into typical OS shell interaction. (I.e. an OS shell is used instead of a Jupyter notebook.)

There are several consequences of this approach:

• Multiple LLMs and LLM providers can be used
• The chat messages can use the provided by the package "LLM::Prompts", [AAp2]:
  • Prompts collection
  • Prompt spec DSL and related prompt expansion
• Easy access to OS shell functionalities

Remark: The package "LLM::DWIM", [BDp1], is very similar in spirit to "Chatnik". "LLM::DWIM" does not use prompt expansion, uses only one chat object, and, although it saves chat history, it does not create chat objects with that history. Both packages are based on the LLM packages "LLM::Functions", [AAp1], and "LLM::Prompts", [AAp2].

Installation

From Zef Ecosystem:

zef install Chatnik

From GitHub:

zef install https://github.com/antononcube/Raku-Chatnik.git

LLM access setup

There are several options for using LLMs with this package:

• Install and run Ollama
  • For the corresponding setup see "WWW::Ollama"
• Run a llamafile / LLaMA model
  • For the corresponding setup see "WWW::LLaMA"
• Have programmatic access to LLMs of service providers like OpenAI or Gemini
  • For the corresponding setup see "WWWW::OpenAI", "WWWW::Gemini", or "WWW::MistralAI"

Basic usage examples

The prompts used in the examples are provided by the Raku package "LLM::Prompts", [AAp2]. Since many of the prompts of that package have dedicated pages at the Wolfram Prompt Repository (WPR) the examples use WPR reference links.

A few turns chat

The script llm-chat is used to create and chat with LLM personas (chat objects):

1. Create and chat with an LLM persona named "yoda1" (using the Yoda chat persona):

llm-chat -i=yoda1 --prompt=@Yoda hi who are you

# Hmmm. Yoda, I am. Jedi Master, wise and old. Guide you, I will. Hmmm. Ask, you must. Answers, I have. Yes, hmmm.

1. Continue the conversation with "yoda1":

llm-chat -i=yoda1 since when do you use a green light saber

# Green, my lightsaber is. Symbol of a Jedi Consular, it is. Wisdom and harmony, it represents. Many years, I have wielded it. Strong in the Force, a Jedi must be, yes. Hmmm. Use the Force, I do, with this green blade. Powerful, it is. Mmm.

Remark: The message input for llm-chat can be given in quotes. For example: llm-chat 'Hi, again!' -i=yoda1.

Remark: The script chatnik can be used instead of llm-chat.

Apply prompt(s) to shell pipeline output

Summarize a file using the prompt "Summarize":

cat README.md | llm-chat --prompt=@Summarize

# Chatnik is a Raku package that provides CLI scripts enabling persistent interaction with multiple Large Language Model (LLM) personas via host OS files, adapting the Jupyter::Chatbook system for shell usage and supporting various LLM providers such as Ollama, Llamafile, OpenAI, Gemini, and MistralAI. It offers tools like `llm-chat` and `llm-chat-meta` for managing chat objects, integrates prompt management through the LLM::Prompts package, and includes features for advanced usage, customization, and planned enhancements like image handling and unit tests. The package's architecture involves persistent chat objects stored as files, prompt expansion, and LLM interaction, allowing flexible, multi-persona LLM conversations directly from the OS shell.

Summarize a file and then translate it to another language using the prompt "Translate":

cat README.md | llm-chat --prompt=@Summarize | llm-chat -i=rt --prompt='!Translate|Russian'

# Chatnik — это пакет для Raku, который предоставляет CLI-скрипты для постоянного взаимодействия с несколькими персонажами Больших Языковых Моделей (LLM) с использованием файлов хост-операционной системы, эффективно адаптируя систему чата Jupyter::Chatbook для использования в оболочке и поддерживая различных провайдеров LLM, таких как Ollama, Llamafile, OpenAI, Gemini и MistralAI. В него входят инструменты, такие как `llm-chat` для общения с персонажами LLM и `llm-chat-meta` для управления объектами чата, интегрируются коллекции и расширения подсказок через пакет LLM::Prompts, и он позволяет выполнять сложные рабочие процессы, такие как суммирование, перевод, создание интеллект-карт и форматированный вывод. С архитектурной точки зрения Chatnik использует JSON-файлы для сохранения объектов чата с рабочим процессом, включающим разбор подсказок, их расширение, взаимодействие с LLM и хранение на базе файлов ОС, поддерживает настройку, текущие задачи разработки и содержит подробные примеры использования.

Remark: The second llm-chat invocation has to use different chat object identifier because the default chat object, with identifier "NONE", is already primed with the prompt "Summary".

Chat objects management

The CLI script llm-chat-meta can be used to view and manage the chat objects used by "Chatnik". Here is its usage message:

llm-chat-meta --help

# Usage:
#   llm-chat-meta <command> [-i|--id|--chat-id=<Str>] [--all] [--<args>=...] -- Meta processing of persistent LLM-chat objects.
#   
#     <command>                  Command, one of: card, clear, delete, file, first-message, last-message, list, load-llm-personas, message, messages.
#     -i|--id|--chat-id=<Str>    Chat id; ignored if --all is specified. [default: 'NONE']
#     --all                      Whether to apply the command to all chat objects or not. [default: False]
#     --<args>=...               Additional, optional arguments for the commands: clear, list, message, messages.

List all chat objects ("chats" and "personas" are synonyms to "list"):

llm-chat-meta list --format=json

# [
#   {
#     "messages": 4,
#     "chat-id": "NONE",
#     "llm-configuration": {
#       "model": "gpt-4.1-mini",
#       "name": "chatgpt"
#     },
#     "context": "Summarize the following text using exactly 3 sentences. Do not add details or editorialize.\n\nThe text to summarize is:"
#   },
#   {
#     "llm-configuration": {
#       "model": "gpt-4.1-mini",
#       "name": "chatgpt"
#     },
#     "messages": 4,
#     "context": "You are Yoda. \nRespond to ALL inputs in the voice of Yoda from Star Wars. \nBe sure to ALWAYS use his distinctive style and syntax. Vary sentence length.",
#     "chat-id": "yoda1"
#   },
#   {
#     "chat-id": "rt",
#     "llm-configuration": {
#       "model": "gpt-4.1-mini",
#       "name": "chatgpt"
#     },
#     "context": "Translate the following text into Russian. Respond with only the translated text. Do not include any explanation or summary.\n",
#     "messages": 2
#   }
# ]

Here we see the messages of "yoda1":

llm-chat-meta messages -i yoda1

# 0 : {
#   "role": "user",
#   "timestamp": "2026-04-28T12:08:15.304329-04:00",
#   "content": "hi who are you"
# }
# 1 : {
#   "content": "Hmmm. Yoda, I am. Jedi Master, wise and old. Guide you, I will. Hmmm. Ask, you must. Answers, I have. Yes, hmmm.",
#   "timestamp": "2026-04-28T12:08:18.483507-04:00",
#   "role": "assistant"
# }
# 2 : {
#   "role": "user",
#   "content": "since when do you use a green light saber",
#   "timestamp": "2026-04-28T12:08:18.890280-04:00"
# }
# 3 : {
#   "timestamp": "2026-04-28T12:08:20.596240-04:00",
#   "role": "assistant",
#   "content": "Green, my lightsaber is. Symbol of a Jedi Consular, it is. Wisdom and harmony, it represents. Many years, I have wielded it. Strong in the Force, a Jedi must be, yes. Hmmm. Use the Force, I do, with this green blade. Powerful, it is. Mmm."
# }

Here we clear the messages:

llm-chat-meta clear -i yoda1

# Cleared the messages of chat object yoda1.

Remark: Calling the script chatnik with the command meta has the same effect as llm-chat-meta. For example, chatnik meta clear -i yoda1 can be used instead of the previous command.

Advanced usage examples

Asking for a result in specific format

llm-chat -i=beta --model=ollama::gemma3:12b 'What are the populations of the Brazilian states? #NothingElse|"JSON data frame, one line per row/dictionary"'

# ```json
# [
#   {"State": "Acre", "Population": 877759},
#   {"State": "Alagoas", "Population": 3432742},
#   {"State": "Amapá", "Population": 857009},
#   {"State": "Amazonas", "Population": 4291852},
#   {"State": "Bahia", "Population": 14743493},
#   {"State": "Ceará", "Population": 9187103},
#   {"State": "Distrito Federal", "Population": 3462537},
#   {"State": "Espírito Santo", "Population": 3967567},
#   {"State": "Goiás", "Population": 7092217},
#   {"State": "Maranhão", "Population": 7016274},
#   {"State": "Mato Grosso", "Population": 3527230},
#   {"State": "Mato Grosso do Sul", "Population": 2740045},
#   {"State": "Minas Gerais", "Population": 21523232},
#   {"State": "Pará", "Population": 8690780},
#   {"State": "Paraíba", "Population": 4051727},
#   {"State": "Paraná", "Population": 11592104},
#   {"State": "Pernambuco", "Population": 9671074},
#   {"State": "Piauí", "Population": 6572551},
#   {"State": "Rio de Janeiro", "Population": 17472043},
#   {"State": "Rio Grande do Norte", "Population": 3535739},
#   {"State": "Rio Grande do Sul", "Population": 11437067},
#   {"State": "Rondônia", "Population": 1150335},
#   {"State": "Roraima", "Population": 531536},
#   {"State": "Santa Catarina", "Population": 7149582},
#   {"State": "São Paulo", "Population": 46287138},
#   {"State": "Sergipe", "Population": 2323117},
#   {"State": "Tocantins", "Population": 1573691}
# ]
# ```

Make a request, echo, and place in clipboard

llm-chat -i=unix '@CodeWriterX|Shell macOS list of files echo the result and copy to clipboard.' | tee /dev/tty | pbcopy

#  ls | tee >(pbcopy) 

Remark: Instead of ... | tee /dev/tty | pbcopy the pipeline command ... | tee >(pbcopy) can be also used.

Make a mind-map of a file

Consider the task of making an (LLM derived) mind map over a certain document. (Say, this REDME.) There are several ways to do that.

1

1. Put file's content to be the positional input argument
2. Use the prompt "MermaidDiagram" in --prompt

llm-chat -i=mmd "$(cat README.md)" --model=ollama::gemma4:26b --prompt=@MermaidDiagram

2

1. Put file's content to be the positional input argument
2. Expand the prompt "manually" via llm-prompt provided by "LLM::Prompts", [AAp2]

llm-chat -i=mmd "$(cat README.md)" --model=ollama::gemma4:26b --prompt="$(llm-prompt 'MermaidDiagram'  below)"

Remark: This example shows another computation result can be used as a prompt. I.e. no need to rely on the automatic prompt expansion.

3

1. Give the prompt "MermaidDiagram" as input
2. Put file's content to be the value of --prompt
   • Put additional prompting for further interaction

llm-chat -i=mmd @MermaidDiagram --model=ollama::gemma4:26b --prompt="FOCUS TEXT START:: $(cat README.md) ::END OF FOCUS TEXT. If it is not clear which text to use, use FOCUS TEXT."

This command allows to do further tasks with the file content as context. For example:

llm-chat -i=mmd '!ThinkingHatsFeedback'

Result

The commands above produce results similar to this diagram:

mindmap
  root("Chatnik")
    Purpose
      Raku package
      CLI for LLM personas
      Persistent interaction via OS files
    Features
      Multiple LLM providers
      LLM Prompts integration
      OS shell access
    LLM Access
      Ollama
      Llamafile
      Service Providers
        OpenAI
        Gemini
        MistralAI
    Scripts
      llm-chat
      llm-chat-meta
        List chats
        Manage messages
        Delete chats
    Installation
      Zef Ecosystem
      GitHub

Render Markdown results with dedicated programs

Get feedback on a text with the prompt "ThinkingHatsFeedback":

cat README.md | llm-chat -i=th --prompt="$(llm-prompt ThinkingHatsFeedback 'the TEXT is GIVEN BELOW.' --format=Markdown)" --model=ollama::gemma4:26b 

Remark: By default the prompt "ThinkingHatsFeedback" gives the hat-feedback table in JSON format. (Currently) the prompt expansion does not handle named parameters, hence, llm-prompt is used to specify the Markdown format for that table.

Get the LLM (chat object) answer -- via llm-chat-meta -- put into a temporary file and "system open" that file:

tmpfile="$TMPDIR/llmans.md"; llm-chat-meta -i=th last-message > "$tmpfile"; open "$tmpfile"

The command above works on macOS. On Linux instead of explicitly creating a file in the temporary dictory, the argument --suffix can be passed to mktemp. For example:

tmpfile=$(mktemp --suffix=".md"); llm-chat-meta -i=th last-message > "$tmpfile"; open "$tmpfile"

Tabulate the LLM personas summary

If the text browser w3m and the Raku package "Data::Translators" are installed, the following pipeline can be used to tabulate the summary the LLM personas:

llm-chat-meta list --format=json | data-translation | w3m -T text/html -dump -cols 120

# ┌────────┬───────┬────────────────────┬───────────────────────────────────────────────────────────────────────────────┐
# │messages│chat-id│ llm-configuration  │                                    context                                    │
# ├────────┼───────┼────────────────────┼───────────────────────────────────────────────────────────────────────────────┤
# │        │       │┌─────┬────────────┐│                                                                               │
# │        │       ││name │chatgpt     ││Translate the following text into Russian. Respond with only the translated    │
# │2       │rt     │├─────┼────────────┤│text. Do not include any explanation or summary.                               │
# │        │       ││model│gpt-4.1-mini││                                                                               │
# │        │       │└─────┴────────────┘│                                                                               │
# ├────────┼───────┼────────────────────┼───────────────────────────────────────────────────────────────────────────────┤
# │        │       │┌─────┬────────────┐│                                                                               │
# │        │       ││model│gpt-4.1-mini││You are Yoda. Respond to ALL inputs in the voice of Yoda from Star Wars. Be    │
# │0       │yoda1  │├─────┼────────────┤│sure to ALWAYS use his distinctive style and syntax. Vary sentence length.     │
# │        │       ││name │chatgpt     ││                                                                               │
# │        │       │└─────┴────────────┘│                                                                               │
# ├────────┼───────┼────────────────────┼───────────────────────────────────────────────────────────────────────────────┤
# │        │       │┌─────┬────────────┐│                                                                               │
# │        │       ││model│gpt-4.1-mini││Summarize the following text using exactly 3 sentences. Do not add details or  │
# │4       │NONE   │├─────┼────────────┤│editorialize. The text to summarize is:                                        │
# │        │       ││name │chatgpt     ││                                                                               │
# │        │       │└─────┴────────────┘│                                                                               │
# ├────────┼───────┼────────────────────┼───────────────────────────────────────────────────────────────────────────────┤
# │        │       │┌──────┬───────────┐│                                                                               │
# │        │       ││ name │ollama     ││                                                                               │
# │2       │beta   │├──────┼───────────┤│                                                                               │
# │        │       ││model │gemma3:12b ││                                                                               │
# │        │       │└──────┴───────────┘│                                                                               │
# └────────┴───────┴────────────────────┴───────────────────────────────────────────────────────────────────────────────┘

Customization

Default model

Default model can be specified with the env variable CHATNIK_DEFAULT_MODEL. For example:

export CHATNIK_DEFAULT_MODEL=ollama::gemma4:26b

Remove with unset CHATNIK_DEFAULT_MODEL.

Pre-defined LLM personas

Use defined LLM personas are specified with JSON file with a content like this:

[
    {
	"chat-id": "raku",
	"conf": "ChatGPT",
	"prompt": "@CodeWriterX|Raku",
	"model": "gpt-4o",
	"max-tokens": 4096,
	"temperature": 0.4
    }
]

(See such a file here.)

The LLM personas JSON file can be specified with the OS environmental variables CHATNIK_LLM_PERSONAS_CONF or RAKU_CHATBOOK_LLM_PERSONAS_CONF -- the former has precedence over the latter.

To load the predefined LLM personas use the command:

llm-chat-meta load-llm-personas

Implementation details

Architectural design

Here is a flowchart that describes the interaction between the host Operating System and chat objects database:

flowchart LR
    OpenAI{{OpenAI}}
    Gemini{{Gemini}}
    Ollama{{Ollama}}
    LLMFunc[[LLM::Functions]]
    LLMProm[[LLM::Prompts]]
    CODBOS[(Chat objects<br>file)]
    CODB[(Chat objects)]
    PDB[(Prompts)]
    CCommand[/Chat command/]
    CCommandOutput[/Chat result/]
    CIDQ{Chat ID<br>specified?}
    CIDEQ{Chat ID<br>exists in DB?}
    IngestCODB[Chat objects file<br>ingestion]
    UpdateCODB[Chat objects file<br>update]
    RECO[Retrieve existing<br>chat object]
    COEval[Message<br>evaluation]
    PromParse[Prompt<br>DSL spec parsing]
    KPFQ{Known<br>prompts<br>found?}
    PromExp[Prompt<br>expansion]
    CNCO[Create new<br>chat object]
    CIDNone["Assume chat ID<br>is 'NONE'"] 
    subgraph "OS Shell"    
        CCommand
        CCommandOutput
    end
    subgraph OS file system
        CODBOS
    end
    subgraph PromptProc[Prompt processing]
        PDB
        LLMProm
        PromParse
        KPFQ
        PromExp 
    end
    subgraph LLMInteract[LLM interaction]
      COEval
      LLMFunc
      Gemini
      OpenAI
      Ollama
    end
    subgraph Chatnik backend
        IngestCODB
        CODB
        CIDQ
        CIDEQ
        CIDNone
        RECO
        CNCO
        UpdateCODB
        PromptProc
        LLMInteract
    end
    CCommand --> IngestCODB
    CODBOS -.-> IngestCODB 
    UpdateCODB -.-> CODBOS 
    IngestCODB -.-> CODB
    IngestCODB --> CIDQ
    CIDQ --> |yes| CIDEQ
    CIDEQ --> |yes| RECO
    RECO --> PromParse
    COEval --> CCommandOutput
    CIDEQ -.- CODB
    CIDEQ --> |no| CNCO
    LLMFunc -.- CNCO -.- CODB
    CNCO --> PromParse --> KPFQ
    KPFQ --> |yes| PromExp
    KPFQ --> |no| COEval
    PromParse -.- LLMProm 
    PromExp -.- LLMProm
    PromExp --> COEval 
    LLMProm -.- PDB
    CIDQ --> |no| CIDNone
    CIDNone --> CIDEQ
    COEval -.- LLMFunc
    COEval --> UpdateCODB
    LLMFunc <-.-> OpenAI
    LLMFunc <-.-> Gemini
    LLMFunc <-.-> Ollama

    style PromptProc fill:DimGray,stroke:#333,stroke-width:2px
    style LLMInteract fill:DimGray,stroke:#333,stroke-width:2px

Here is the corresponding UML Sequence diagram:

sequenceDiagram
    participant CCommand as Chat command
    participant IngestCODB as Chat objects file ingestion
    participant CODBOS as Chat objects file
    participant CODB as Chat objects
    participant CIDQ as Chat ID specified?
    participant CIDEQ as Chat ID exists in DB?
    participant RECO as Retrieve existing chat object
    participant PromParse as Prompt DSL spec parsing
    participant KPFQ as Known prompts found?
    participant PromExp as Prompt expansion
    participant COEval as Message evaluation
    participant CCommandOutput as Chat result
    participant CNCO as Create new chat object
    participant CIDNone as Assume chat ID is NONE
    participant UpdateCODB as Chat objects file update
    participant LLMFunc as LLM Functions
    participant LLMProm as LLM Prompts

    CCommand->>IngestCODB: Chat command
    CODBOS--)IngestCODB: Chat objects file
    IngestCODB--)CODB: Chat objects
    IngestCODB->>CIDQ: Chat ID specified?
    CIDQ-->>CIDEQ: Yes
    CIDQ-->>CIDNone: No
    CIDNone->>CIDEQ: Assume chat ID is NONE
    CIDEQ-->>RECO: Yes
    CIDEQ-->>CNCO: No
    CIDEQ--)CODB: Chat objects
    RECO->>PromParse: Prompt DSL spec parsing
    PromParse--)LLMProm: LLM Prompts
    CNCO--)LLMFunc: LLM Functions
    CNCO--)CODB: Chat objects
    CNCO->>PromParse: Prompt DSL spec parsing
    PromParse->>KPFQ: Known prompts found?
    KPFQ-->>PromExp: Yes
    KPFQ-->>COEval: No
    PromExp--)LLMProm: LLM Prompts
    PromExp->>COEval: Message evaluation
    COEval--)LLMFunc: LLM evaluator invocation
    LLMFunc--)COEval: Evaluation result
    COEval->>UpdateCODB: Chat objects file update
    COEval->>CCommandOutput: Chat result

Persistent chat objects

Using a JSON file for keeping the chat objects database is a fairly straightforward idea. Efficiency considerations for "using the OS to manage the database" are probably can not that important because LLMs invocation is (much) slower in comparison.

Remark: The following quote is attributed to Ken Thompson about UNIX:

We have persistent objects, they're called files.

TODO

• TODO Implementation
  • DONE Chats DB export
  • DONE Chats DB import
  • DONE LLM persona creation
  • DONE LLM persona repeated interaction
  • DONE CLI llm-chat
    • DONE Simple: $input & *%args
    • DONE Multi-word: @words & *%args
    • DONE From pipeline
    • CANCELED Format?
  • DONE CLI llm-chat-meta
    • DONE Commands reaction
    • DONE View messages for an id
    • DONE Clear messages for an id
    • DONE Delete chat for an id
    • DONE View all chats
    • DONE Delete all chats
    • DONE Clear message for an id by range
    • DONE Take message for an id by index
    • DONE Take last message for an id
    • DONE Load LLM personas in the JSON file used for initialization by "Jupyter::Chatbook"
  • TODO Handling images
    • TODO Handle image path specified with the option --image
    • TODO Recognizing the input is or has an image file path
    • TODO Handle output of image generation models
    • TODO Use cases:
      • TODO Image generation from a textual description
      • TODO Extract and translate textual content of an image
      • TODO Make another image from a given image
• TODO Unit tests
  • DONE Export & import
  • DONE Main workflow
  • DONE Persona repeated interaction
  • DONE Persona creation
  • TODO CLI tests
• TODO Documentation
  • DONE Flowchart & sequence diagram
  • DONE Usage examples
    • DONE Basic examples
    • DONE Advanced examples
    • DONE Management (meta) examples
  • TODO Demo video

References

Articles, blog posts

[AA1] Anton Antonov, "Chatnik: LLM Host in the Shell — Part 1: First Examples & Design Principles", (2026), RakuForPrediction at WordPress.

Packages

[AAp1] Anton Antonov, LLM::Functions, Raku package, (2023-2026), GitHub/antononcube.

[AAp2] Anton Antonov, LLM::Prompts, Raku package, (2023-2025), GitHub/antononcube.

[AAp3] Anton Antonov, Jupyter::Chatbook, Raku package, (2023-2026), GitHub/antononcube.

[BDp1] Brian Duggan, LLM::DWIM, Raku package, (2024-2025), GitHub/bduggan.

[JSp1] Jonathan Stowe, XDG::BaseDirectory, Raku package, (2016-2026), GitHub/jonathanstowe.