AI Tools

AI is where autonomy gets lost fastest — and where the difference between control and dependence is one architectural decision. Every tool below can run without sending your data to someone else’s server.

Most AI tools exist in two modes: hosted (convenient, dependent) and self-hosted (requires effort, sovereign). TAS scores both. A low score is not a criticism — it is a measurement of what you trade for convenience.

Training a model costs millions. Companies need to recoup that investment — hosted APIs are how. That’s not a flaw, it’s economics. But you should know exactly what you’re paying: money, or control.

Unlike most software categories, AI gives you a real choice. The same model that runs behind a $20/month API can often run on your own hardware for free. TAS makes that difference visible.

Choose Your Level

3 levels of commitment · 15+ tool recommendations

Use Local Models

Install Ollama or LM Studio. Run open-weight models on your own machine. Your prompts never leave your device. No account needed.

Self-Host the Stack

Run Ollama + Open WebUI on your server. Add Whisper for speech, SearXNG for search. Multi-user, full control, ChatGPT-like experience.

Full Sovereign AI

Air-gapped inference. No internet. Models on local storage, custom fine-tuning, own hardware. Maximum autonomy — requires GPU investment.

The Two-Mode Problem

Almost every AI product exists in two configurations — and the TAS score changes dramatically between them:

Tool	Self-hosted	Hosted
MiroThinker	A3/T2 (open weights, run locally)	A1/T1 (hosted app, account required)
ChatGPT	—	A0/T0 (cloud-only, proprietary)
Whisper	A3/T2 (run locally)	A0/T0 (OpenAI API)
Stable Diffusion	A3/T2 (ComfyUI local)	A0/T0 (cloud services)

This is not a bug — it’s the core insight. The same intelligence, radically different autonomy.

Inference Engines

The runtime that loads and serves models on your hardware.

Runs open-weight models locally (Llama, Qwen, Mistral)

One-command install, OpenAI-compatible API. The Docker of LLMs. Pull a model, run it, done.

→ Local alternative to: cloud inference APIs

The foundation. C++ inference engine that runs on CPU and GPU. Maximum control, minimum abstraction.

→ Local alternative to: hosted inference APIs

Production-grade GPU serving with PagedAttention. High throughput for multi-user deployments.

→ Local alternative to: the OpenAI API

Drop-in OpenAI API replacement. Supports text, image, audio generation. One binary, no GPU required.

Chat Interfaces

Give your local models a usable frontend.

→ Local alternative to: the ChatGPT interface

Self-hosted ChatGPT-like UI. Connects to Ollama or any OpenAI-compatible API. Multi-user, RAG, web search built in.

→ Local alternative to: ChatGPT desktop

Desktop app for running local models. Beautiful UI, model discovery, chat. Free but closed-source.

→ Local alternative to: ChatGPT desktop

Open-source desktop AI. Runs models locally, no internet needed. AGPL-3.0 licensed.

→ Local alternative to: ChatGPT

Desktop app from Nomic. Local chat, document Q&A. Optimized for consumer hardware.

Reasoning & Research Models

Open-weight models designed for deep reasoning and agentic research.

→ Local alternative to: hosted deep-research agents

A3/T2 (self-hosted) · A1/T1 (hosted)

30B/235B reasoning model from MiroMind. Verification-centric architecture. Open weights on HuggingFace, run via vLLM/SGLang. Hosted app exists but requires account.

DeepSeek-R1 family

→ Local alternative to: cloud reasoning APIs

A3 · T1–T2 (depends on variant) · hosted API A1/T1

Open-weight reasoning, chain-of-thought visible. Three scored variants: full 671B (MIT, T2), Qwen-based distills (Apache base, T2), Llama-based distills (contested license, T1). Distills run on consumer hardware via Ollama.

Speech & Audio

Local speech-to-text and text-to-speech.

→ Local alternative to: Google Speech-to-Text, Otter.ai

OpenAI's Whisper model, ported to C++. Runs on CPU. Transcribe audio locally — no data leaves your machine.

→ Local alternative to: Google TTS

Fast local text-to-speech. Multiple languages, natural voices. Runs on Raspberry Pi.

Image Generation

Local image generation and editing.

→ Local alternative to: Midjourney, DALL-E

Node-based UI for Stable Diffusion and other image models. Full local workflow — generate, edit, upscale. No cloud.

→ Local alternative to: Midjourney, DALL-E

Web UI for Stable Diffusion. The original self-hosted image generation interface. Huge extension ecosystem.

Hosted Services (for comparison)

These are the services that self-hosted AI replaces. Included for contrast — not as recommendations.

ChatGPT / OpenAI API

Cloud-only. Proprietary models. All prompts and data processed on OpenAI servers. Cannot self-host. Cannot export model.

Cloud-only. Proprietary models by Anthropic. No self-hosted option. Data processed on Anthropic/cloud infrastructure.

Cloud-only. Google infrastructure. Gemma models are open-weight (A3/T2 when self-hosted), but Gemini API is not.

Cloud-only image generation. Discord-based interface. No self-hosted option. No open model.

What’s Next

Set up local AI? Here’s where to go deeper:

Recommended Server Stack

→ build the full backend

→ deploy complete stacks

Audit Your Setup

→ score any technology

→ 150 technologies evaluated

Recommended Stack → Recipes → Full catalog → Suggest a technology →