Antri CLI

A unified, multi-surface agentic coding ecosystem featuring dialectic debates, live interactive artifacts, and lifelong thinking profiles.

Comment

Inspiration

As developers, we spend hours bouncing between browser chat tabs, terminal windows, code editors, and mobile notes. Most AI coding assistants are still trapped in single-turn text boxes—they give you raw markdown code blocks you have to copy-paste, forget your architecture preferences after a few messages, and cannot reason beyond shallow first guesses.

I asked myself: What if an AI companion wasn’t just a chatbot, but a true multi-surface agentic
ecosystem?

I wanted a tool that lives in the terminal where I work, synchronizes with a native mobile client when I’m away from my desk, tests its own assumptions through background dialectic debates, and
doesn't just spit out code—but generates live, interactive runnable applications and visual mind maps on the fly. That vision became ANTRI Code.
──────
## What it does

ANTRI is an autonomous, multi-surface agentic coding and reasoning ecosystem that connects a
Terminal CLI, a Desktop Control Plane, and a Native Flutter Mobile App.

Key capabilities include:

• Autonomous Dialectic Reasoning & Goal Loops (/debate, /goal): When asked complex architectural questions, ANTRI silently runs a multi-agent adversarial debate (Thesis → Antithesis → Empirical Verification → Consensus) behind the scenes, delivering battle-tested synthesized solutions.
• Claude-Style Live Interactive Artifacts (/view): Instead of dumping raw HTML/JS code into chat, ANTRI renders full Multi-Page Single-Page Applications (3–10 navigable tabs, interactive timers, live checklists, calculators) directly inside the app with zero code exposure.
• Markmap Collapsible Mind Maps (/mindmap): Renders reference-grade hierarchical visual trees
with 60fps pan/zoom, interactive subtree collapse/expand nodes, and porcelain light / poster dark themes.
• Persistent Thinking Profiles & Memory (/profile, /notes): Learns your coding conventions,
preferred paradigms, and project rules automatically, preserving them in structured Markdown
across sessions and devices.
• Multi-Provider Engine: Seamlessly switch between Gemini 1.5/2.0, Claude, GPT-4o, DeepSeek,
NVIDIA NIM, and local Ollama models.
──────
## How we built it

We built ANTRI with a modular, performance-obsessed architecture:

• Core CLI Engine: Built in TypeScript (NodeNext ESM) using Commander and prompt_toolkit-style
ANSI box renderers for real-time floating / autocomplete and streaming token pipelines.
• Mobile Client: Developed in Flutter / Dart (antri_flutter) to deliver 60fps gesture handling,
multi-touch pinch-to-zoom for mind maps, and a native WebView sandbox for live SPA artifacts.
• Interactive Artifact Engine: Powered by Markmap + D3 and a custom Aurora Glassmorphism CSS/JS
runtime that embeds interactive web apps and flowcharts directly inside chat messages.
• Workspace Tool Runner: Sandboxed file system operations (edit_file, grep_search, find_files,
execute_python, git_diff) with automated self-debugging loops.
• Distribution: Packaged and published production-ready on the official npm registry as antri_cli https://www.npmjs.com/package/antri_cli.
──────
## Challenges we ran into

  1. Multi-Platform Visual Parity: Ensuring that complex SVG mind maps and multi-page HTML
    artifacts rendered with identical 60fps smoothness across a terminal emulator, desktop browser,
    and mobile Android WebView took significant tuning.
  2. Preventing LLM Tool Misuse on Small Models: When testing smaller open-source models (e.g.
    Llama 3.1 8B on NVIDIA NIM), we noticed models hallucinated run_command with shell echo
    statements for simple greetings. We built an intelligent echo-interception layer that extracts
    the speech, bypasses permission gates, and renders it cleanly to the CLI.
  3. Low-Latency Background Dialectics: Running multi-persona debates and 3-stage goal loops
    without slowing down user experience required careful stream orchestration and synthesized
    progress badges.
    ──────
    ## Accomplishments that we're proud of

• True Multi-Surface Fluidity: Switching seamlessly from a terminal CLI on PC to a native Flutter app on mobile while maintaining active chat sessions, cognitive memory, and artifacts.
• Live Runnable SPAs in Chat: Generating full 7-day interactive workout dashboards, calculators, and trackers that users can interact with immediately without ever seeing raw boilerplate.
• Production npm Package & Mobile APK: Going from concept to a published npm package with 47/47
passing automated unit tests and a release-ready Android APK.
## What we learned

• Agentic thinking beats raw parameter size: A smaller model guided by a structured dialectic
consensus loop (Thesis vs. Antithesis vs. Verifier) consistently out-reasons a larger model
responding in a single shot.
• Visual execution creates delight: Developers love the terminal, but complex plans and
architectures are far more intuitive when rendered as collapsible trees and interactive tabs.
──────
## What's next for Antri CLI

• Multimodal Camera / Screenshot Bug Triage: Snap a photo of a broken terminal or UI mockup in
Flutter mobile and have ANTRI automatically reproduce and patch the code in your workspace.
• Collaborative Swarm Workspaces: Allowing team members to share active Thinking Profiles and
live Artifact sandboxes in real-time.
• Google Search Grounding & Cloud Vertex Deployment: Deepening Gemini-native search grounding for instant library updates and zero-config cloud workspaces.

Share this project:

Updates

Submission history