GPTCLI

GPTCLI is a CLI client written entirely in Python for accessing the LLM of your choice without the need for Web or Desktop apps. All data is encrypted at rest using AES-256-GCM.

How to run it

Summary:

• PyPI:
  • Install with pip install dbc-gptcli.
  • Run gptcli [mistral|openai] [chat|se|ocr].
• Docker:
  • Pull with docker pull deathbychocolate/gptcli:latest.
  • Start a container with docker run --rm -it --entrypoint /bin/bash deathbychocolate/gptcli:latest.
  • Run gptcli [mistral|openai] [chat|se|ocr].

For more info on usage, check the builtin help docs with:

• gptcli -h
• gptcli [mistral|openai] [chat|se|ocr] -h.

Command tree

gptcli [--no-cache] [--loglevel LEVEL]
├── all
│   ├── encrypt       # Encrypt all cleartext files
│   ├── decrypt       # Decrypt all encrypted files
│   ├── rekey         # Re-encrypt with a new passphrase
│   └── nuke          # Permanently delete all gptcli data
├── mistral
│   ├── chat          # Multi-turn conversation
│   ├── se            # Single exchange
│   ├── ocr           # Document to Markdown conversion
│   └── search
│       ├── chat      # Full-text search over chat history
│       └── ocr       # Full-text search over OCR history
└── openai
    ├── chat          # Multi-turn conversation
    ├── se            # Single exchange
    └── search
        └── chat      # Full-text search over chat history

How to get an API key

You need valid API keys to communicate with the AI models.

For OpenAI:

• Create an OpenAI account here: https://chat.openai.com/
• Generate an OpenAI API key here: https://platform.openai.com/api-keys

For Mistral AI:

• Create a Mistral AI account here: https://chat.mistral.ai/chat
• Generate a Mistral AI API key here: https://console.mistral.ai/api-keys

How it works

The project uses the API of LLM providers to perform chat completions. It does so by sending message objects converted to JSON payloads and sent over HTTPS POST requests.

GPTCLI facilitates access to 2 LLM providers, Mistral AI and OpenAI. Each provider offers modes to communicate with the LLM of your choosing: Chat, Single-Exchange, and OCR (Mistral only).

Modes

Chat

Chat mode allows the user to have a conversation that is similar to ChatGPT by creating a MESSAGE-REPLY thread. For example, you say hello:

You can have a conversation multiline conversations:

You can load the last conversation you had with the LLM provider (OpenAI, Mistral):

And if you want to know about in-chat commands, you can view them by asking for help:

Chat mode also automatically:

• Stores chats locally as oneline json files via the --store and --no-store flags.
• Uses previously sent messages as context via the --context and --no-context flags.
• Loads the provider's API key; you may overwrite this behaviour by providing a different key with the --key flag.

Single-Exchange (SE)

Single-Exchange is functionally similar to chat mode, but it only allows one exchange of messages to happen (1 message sent from client-side and 1 response message from server-side) and then exit. This encourages loading all the context and instructions in one message. It is also more suitable for automating multiple calls to the API with different payloads, and flags. This mode will show you output similar to the following:

This mode does not store chats locally. It is expected the user implements their own solution via piping or similar.

OCR (Optical Character Recognition)

OCR mode converts documents (PDFs, images) into Markdown text. Currently available for Mistral AI only. It accepts local filepaths and/or URLs as arguments, or a batch of documents via --filelist. By default, results are saved as Markdown files in the current directory.

OCR mode also automatically:

• Stores OCR results locally via the --store and --no-store flags.
• Saves converted Markdown files to the current directory; you may change this with --output-dir or disable it with --no-output-dir.
• Supports batch processing from a file of paths/URLs via the --filelist flag.
• Displays the Markdown result to stdout via the --display flag.
• Displays the most recent OCR session from storage via the --display-last flag.
• Excludes images from the OCR response via the --no-images flag, returning only Markdown text.

Search (Full-Text Search)

Search mode provides an interactive TUI for full-text search over locally stored history. It is available for both providers under chat, and for Mistral AI also under ocr.

gptcli mistral search chat   # Search Mistral chat history
gptcli openai search chat    # Search OpenAI chat history
gptcli mistral search ocr    # Search Mistral OCR history

Type to filter results in real time. Chat search navigation and actions:

OCR search navigation and actions:

For OCR search, the output directory for Ctrl+W can be set with --output-dir (defaults to .).

Encryption

GPTCLI encrypts all data at rest using AES-256-GCM with scrypt key derivation. On first run, you are prompted to create a passphrase (16 characters minimum). The derived encryption key is cached for 12 hours using a wrapping key in volatile storage, so you don't need to re-enter your passphrase on every invocation.

You can manage encryption across all providers with:

• gptcli all encrypt — Encrypt all cleartext files.
• gptcli all decrypt — Decrypt all encrypted files.
• gptcli all rekey — Re-encrypt all files with a new passphrase.

Use the --no-cache flag to disable key caching and prompt for the passphrase every time.

Features

Implemented

• Send text based messages to Mistral AI API.
• Send text based messages to OpenAI API.
• Store API keys locally.
• Allow context retention for chats with all providers.
• Allow streaming of text based messages for all providers.
• Allow storage of chats locally for all providers.
• Allow loading of chats from local storage as context for all providers.
• Add in-chat commands.
• Add multiline option for chat mode.
• Add spinner animation for chat mode.
• Add OCR as a new mode.
• Send OCR queries for images and PDF documents to Mistral AI API.
• Allow storage of OCR results locally.
• Add encryption at rest for all locally stored data.
• Add key caching with automatic expiry.
• Add passphrase rekeying.
• Add FTS for chats in storage.
• Add FTS for OCR results in storage.
• Add nuke command to permanently delete all gptcli data.

In Development

• Send OCR queries for images and PDF documents to OpenAI API.
• Add role-based messages for Mistral AI: user system assistant developer tool function
• Add role-based messages for OpenAI: user system assistant developer tool function

Lexicon

How GPTCLI is different from other clients

• GPTCLI does not use any software developed by OpenAI or Mistral AI, except for counting tokens.
• GPTCLI prioritizes features that make the CLI useful and easy to use.
• GPTCLI aims to eventually have all the features of its WebApp counterparts in the terminal.