lana-gemcli 1.0.0

a simple yet useful cli for interacting with google gemini

lana

a simple semi-agentic chatbot interface for gemini built around the terminal.

context

i wanted aistudio in the terminal. but i kept going.

how ai was used during the creation of this project

• used to help understand docs
• used to implement a prototype solution to one complicated problem that i then studied/understood and replaced w/ my own solution
• the project itself is merely a client for AI APIs

getting started

this project is on pypi! you just have to install python and

pip install lana-gemcli

roadmap/to-do

core

• basic gemini api interface
• conversation
• basic search tool calls
• file manipulation and code execution tools (make these not suck)
• sysprompt encouraging heavy tool usage
• asyncification
• switch to selenium

user interaction

• multimodal inputs
• multi-line user prompts
• user image/video/etc input
  • argparse
  • mime type detection
  • inserting files into prompts after the first, TBD alongside UI improvements
• user-defined system prompt
• improve ui and ux
  • rich formatting
  • configuration/settings interface
  • saved chats interface
• easy installation
• tba

tools and data retrieval

• selenium improvements
  • switch to selenium headless
  • click on element with css selector
  • type text with css selector
  • take screenshot
• persistent chats
  • context-to-file serialisation
  • file-to-context deserialisation
• tba

agentic functionality

• persistence
  • persistent chats, see "tools and data retrieval"
  • "notebooks"
    • basically just markdown files that contain certain kinds of info
    • similar to openclaw's SOUL.md, MEMORY.md, HEARTBEAT.md, etc.
    • model can literally just write to these using the provided tools i think?
• sub-agents
  • sub-agents. the "root" agent should be able to activate these via tool call and they run in the background
  • this would require some kind of multi-threading or idk maybe asyncio has something useful?
  • have the main loop poll the agents on every generation, so we can inject the sub-agent return as a tool return even if the sub-agent tool already returned "sub-agent started" or similar.
  • sub-agents should be able to spawn their own sub-agents (should they be? i'm not sure about this), which means that certain environment variables (system prompt, running/owned sub-agents, etc.) should be tied to an agent. speaking of which, i should probably make this more OOP-like (see: internal codebase changes)
  • post ship
• tba

internal codebase

• tool-calling boilerplate cleanup
• stronger typing
• tba