I Built a Terminal AI Assistant That Never Freezes — Here's the Architecture

Devesh tiwari — Wed, 26 Aug 2026 06:32:18 +0000

I'm an AI Engineering student, and I got tired of alt-tabbing to a browser
every time I needed to ask an LLM something mid-coding session. So I built
term-ai — a fully async, terminal-native AI assistant that runs as a
real installed CLI tool, not a janky script.

Here's what I built, why I made specific architectural decisions, and what
broke along the way.

What It Does

Type ai-agent in any terminal. Get a clean, keyboard-driven chat interface
that renders AI responses as live Markdown — code blocks, headings, lists —
right in your shell. Switch between multiple free AI models mid-conversation
without losing context. That's the whole pitch.

The Stack

Textual — for the async TUI
LangChain — for conversation memory and orchestration
OpenRouter API — for routing prompts to free LLM models
pyproject.toml — packaged as a real CLI tool

Architecture: Three Layers, One Rule

The one rule: nothing blocks the UI thread.

Presentation Layer — Textual

Every network call runs inside a Textual @work worker task. This is the
entire reason the UI never freezes — even on a slow model response, you can
keep scrolling, typing, or switching models while generation runs in the
background.

Responses render as live Markdown. This matters more than it sounds —
getting back a wall of **bold** and code asterisks in a
terminal is genuinely painful.

Orchestration Layer — LangChain

LangChain manages in-memory conversation state. The key design decision here:
memory is not tied to a single model. You can start a task with one model,
switch to another mid-conversation, and the full context carries over. Same
thread, different brain, zero friction.

Routing Layer — OpenRouter

OpenRouter gives you a single API endpoint that routes to 400+ models. For
this project I'm routing exclusively to free-tier models.

The Problem That Almost Broke Everything

I had a hardcoded AVAILABLE_MODELS list:

AVAILABLE_MODELS = [
    ("Gemini 2.5 Pro", "google/gemini-2.5-pro:free"),
    ("DeepSeek R1",    "deepseek/deepseek-r1:free"),
    ("Llama 4 Maverick", "meta-llama/llama-4-maverick:free"),
    # ...
]

One day, every single model started throwing:

The entire top tier had quietly gone paid. The :free suffix in the model
id doesn't mean anything once a provider pulls the model from the free tier
— OpenRouter just updates the pricing fields in their API. The id string
stays the same.

The Permanent Fix

I wrote openrouter_models.py — a self-healing module that fetches
OpenRouter's live model catalog on startup and filters by actual pricing
data, not by the id string:

def _is_free(model: dict) -> bool:
    pricing = model.get("pricing") or {}
    prompt_price = pricing.get("prompt")
    completion_price = pricing.get("completion")

    # Pricing is the source of truth — not the :free suffix.
    # A model can keep a :free-looking id after going paid.
    if prompt_price is not None and completion_price is not None:
        return prompt_price in ("0", 0) and completion_price in ("0", 0)

    return str(model.get("id", "")).endswith(":free")

This runs once on startup, caches for 6 hours locally, and falls back
gracefully if OpenRouter is unreachable. The AVAILABLE_MODELS list now
rebuilds itself every time the cache expires — I will never manually patch
a model slug again.

Installation (The Actual Modern Way)

This is packaged via pyproject.toml, not a shell script:

git clone https://github.com/your-username/ai-terminal-app
cd ai-terminal-app
python -m venv venv && source venv/bin/activate
cp .env.example .env   # add your OpenRouter key
pip install -e .

Then from anywhere:

ai-agent

For a true global install outside any venv:

pipx install .

Keybindings

Key	Action
`Enter`	Submit prompt
`Ctrl+L`	Clear memory
`Ctrl+C`	Quit
Mouse	Scroll, click to navigate

What's Next

🎙️ Voice Engine — local Whisper STT + TTS, fully offline
💾 Persistent Memory — SQLite + LangChain so context survives between sessions

DEV Community: Devesh tiwari