paritok-cli

A cost-optimized interactive coding agent CLI. Routes requests through the paritok proxy for compression, then forwards to NVIDIA's API for AI-powered code assistance.

prompt > build a calculator using typescript
AI     → generates code
CLI    → runs it, captures output
AI     → fixes errors, iterates
CLI    → returns working code

How It Works

┌──────────────┐     POST /v1/chat/completions     ┌──────────┐     NVIDIA API     ┌─────────┐
│ paritok-cli  │ ──── Authorization: NVIDIA key ──→ │ paritok  │ ────────────────→ │ NVIDIA  │
│  (chat TUI)  │     (streaming, no compression)    │  proxy   │                   │  API    │
└──────────────┘                                    └──────────┘                   └─────────┘
       │                                                   ▲
       │ POST https://www.paritok.com/api/compress          │
       │ (compresses content before sending)                │
       └────────────────────────────────────────────────────┘

The CLI sends chat completions through the local paritok proxy (port 8080). The proxy forwards to NVIDIA. Separately, the CLI compresses content via paritok's cloud API before sending to save tokens.

Prerequisites

• Go 1.26+ — to build the CLI
• paritok proxy — the local forwarding proxy
• PARITOK_API_KEY — from paritok.com (for compression)
• NVIDIA_API_KEY — from NVIDIA API (for AI inference)

Installation

1. Build the CLI

git clone https://github.com/LyWilson/paritok-cli.git
cd paritok-cli
go build -o paritok-cli.exe .

2. Install the paritok proxy

pip install paritok

Make sure paritok.exe is on your PATH (check with where paritok).

3. Configure paritok.yaml

Edit paritok.yaml in the project directory:

use_gpu_server: true
gpu_server:
  api_key: "pk_live_<your-paritok-api-key>"

4. Save your API keys

paritok-cli auth pk_live_<your-paritok-key>
paritok-cli auth-nvidia nvapi-<your-nvidia-key>

Keys are stored in ~/.paritok.json with 0600 permissions.

Usage

Quick start

# Terminal 1: start the proxy
paritok-cli proxy start

# Terminal 2: chat
paritok-cli chat

# When done
paritok-cli proxy stop

Commands

Command	What it does
chat	Interactive TUI — type messages, AI streams responses
code <prompt>	Headless coding agent — generates code, runs it, iterates to fix errors
proxy start	Starts the paritok proxy on port 8080
proxy stop	Stops the proxy
auth <key>	Save paritok API key to ~/.paritok.json
auth-nvidia <key>	Save NVIDIA API key to ~/.paritok.json

TUI chat commands

Command	What it does
/help	Show available commands
/quit / /exit	Exit the TUI
/code <prompt>	Enter code iteration mode inside the chat

Coding agent

Headless mode:

paritok-cli code "build a calculator using typescript"

This generates code, saves it to a temp file, runs it (e.g. npx tsx for TypeScript), captures output, and feeds errors back to the AI to fix. Loops up to 5 iterations.

In the TUI:

/code build a calculator in python

Same loop but progress is shown inline in the chat viewport.

Supported languages:

Language	Runner
TypeScript / .ts	npx --yes tsx <file>
JavaScript / .js	node <file>
Python / .py	python <file>
Go / .go	go run <file>
Bash / .sh	bash <file>

Model selection

Default model: openai/gpt-oss-20b

paritok-cli chat --model "meta/llama-3.1-70b-instruct"
paritok-cli code --model "meta/llama-3.1-70b-instruct" "build a calculator"

Configuration

All config is in ~/.paritok.json:

{
  "api_key": "pk_live_...",
  "nvidia_api_key": "nvapi-...",
  "base_url": "http://127.0.0.1:8080"
}

Setting	Env var	Default
api_key	PARITOK_API_KEY	—
nvidia_api_key	NVIDIA_API_KEY	—
base_url	PARITOK_BASE_URL	http://127.0.0.1:8080

Env vars take precedence over the config file.

Built With

• Paritok — Prompt compression for cost-optimized LLM usage
• NVIDIA API — AI inference

License

Apache 2.0