The good stuff, with the catch attached.

The rules. Open-source by default. Paid tools get a nod only when they have earned it in daily use, and the only paid one on that list right now is Anthropic's Claude and Claude Code. That recommendation pays this site exactly nothing, and the zero is the point. Everything below is free to run on hardware you already own.

Open-source AI

Buzz

Flagship

Block's open-source, agent-first workspace where every human and every AI agent is a first-class citizen with its own cryptographic identity, and you can host the whole thing yourself.

Good at: Giving people and agents their own keypair identity, with every message, workflow, and git event written as a signed entry in an append-only, verifiable audit log. Self-hostable on your own infrastructure, so you control the data, the relay, and the agents. Built on the Nostr protocol, which means an agent here is a keypair with its own audit trail, not a bot token some vendor can switch off. Released by Block (Jack Dorsey's company) on July 21, 2026, under Apache 2.0, and pointed straight at proprietary lock-in like Slack and GitHub.
The catch: It is early, and Block says so themselves. It is around version 0.4, the mobile apps are not finished, push notifications are still pending, and standing it up yourself is real engineering (Docker, Rust, Node, and running your own relay). It is not a polished turnkey product yet. And to head off the obvious question: it is not blockchain. The signed-events design is about auditability and identity, not crypto.

github.com/block/buzz · engineering.block.xyz/blog/buzz

Run your own models

Ollama

The easiest on-ramp to running open language models on your own machine. Pull a model, talk to it over a local API.

Good at: One-command model management and a local OpenAI-compatible API that almost every other tool can plug into. Runs on modest hardware with quantized models, and it is one of the most widely used ways people get started with local AI.
The catch: It is a convenience layer over llama.cpp, so you trade some low-level control and tuning for the ease. The defaults are not always the best settings for your hardware, and big models still want real RAM or VRAM that a wrapper cannot conjure.

ollama.com · Pairs with Field Log 001, run your own AI

llama.cpp

The C and C++ engine that made local language models practical on normal hardware. Most of the local-AI stack sits on top of it.

Good at: Efficient inference across a huge range of machines, CPU or GPU, with fine-grained control over quantization and memory. It is the foundation under a lot of friendlier tools, Ollama included, and it has a large active community.
The catch: It is lower level. You deal with model files, quantization formats, and build flags yourself. More power and more setup than a wrapper, which is exactly why most people run it through one.

github.com/ggml-org/llama.cpp · Pairs with Field Log 001

Open WebUI

A self-hosted, ChatGPT-style web front-end for your local models. Makes a model on your own box feel like a real product.

Good at: A polished multi-user chat interface with document chat, model switching, and per-user history, all self-hosted. It works with Ollama and any OpenAI-compatible backend, so it drops onto whatever you are already running.
The catch: It is the face, not the brain. You still need a model backend behind it, and the steady feature sprawl means there is more to configure and more to keep updated than a plain chat box.

github.com/open-webui/open-webui · Pairs with Field Log 001

ComfyUI

A node-graph interface for local image and video generation. You wire the pipeline yourself instead of pushing one button.

Good at: Total control over diffusion pipelines with reproducible, shareable workflows, running open image models on your own GPU. It is deeply extensible, which is why it became the power-user tool of choice for local image work.
The catch: The node graph is a real learning curve, and it is a security footgun: custom nodes and downloaded workflows can run arbitrary code on your machine, so only load ones you trust. It also wants a capable GPU to be pleasant.

github.com/comfyanonymous/ComfyUI

Voice and vision at home

Home Assistant Assist, with Whisper and Piper

Fully local voice control for your smart home. Speech in (Whisper) and speech out (Piper) run on your own hardware, no cloud.

Good at: Private, offline voice commands with no account and no audio leaving the house, wired straight into the local-first Home Assistant stack. You own the whole voice loop.
The catch: Local speech recognition is heavier and less forgiving than the big cloud assistants. Accuracy and response time depend on your hardware, and the setup is more involved than plugging in a smart speaker.

home-assistant.io/voice_control · Pairs with Field Log 004, Home Assistant

Frigate

A self-hosted camera recorder with local AI object detection. It alerts you on people and cars, not every branch that moves.

Good at: Real-time object detection that runs on your own box, no cloud camera subscription, with tight Home Assistant integration. It shines paired with a cheap AI accelerator so detection stays smooth.
The catch: It wants a dedicated setup and, realistically, a hardware accelerator to run well. Getting cameras, zones, and detectors configured is genuine work, not plug-and-play.

frigate.video · Pairs with Field Log 004

Coding and agent platforms

Aider

An AI pair programmer in your terminal that edits your git repo directly and commits as it goes.

Good at: Fast, code-aware edits across a real repository with git baked in, so every change is a reviewable commit. It is model-agnostic, so you can point it at a local model or a hosted one.
The catch: It is only as good as the model behind it, and it will make confident wrong edits, so you are the reviewer every time. The strongest results come from paid hosted models, and local models are noticeably weaker at this kind of work.

aider.chat · Pairs with Field Log 006, build your own AI agent

OpenHands

An open-source autonomous coding agent that can write code, run commands, and drive a browser to finish dev tasks.

Good at: More hands-off, multi-step development than a pair programmer, and it is self-hostable with active development behind it. Good when you want an agent to attempt a whole task, not just the next line.
The catch: Autonomy cuts both ways. It runs real commands and can wander off the rails, so you sandbox it and watch it. Results are uneven and lean heavily on the quality of the underlying model.

github.com/All-Hands-AI/OpenHands · Pairs with Field Log 006

Dify

A self-hostable platform for building LLM apps and agents with a visual builder, RAG, workflows, and tools included.

Good at: Getting from idea to a working LLM app or agent fast, without wiring every piece by hand, and self-hosted so your prompts and data stay yours. A solid on-ramp for people who do not want to start from a blank file.
The catch: The visual-platform convenience buys you less low-level control, and running the full stack yourself is heavier than a single script. You can outgrow its abstractions once your needs get specific.

dify.ai · Pairs with Field Log 006

LangGraph and Deep Agents

A code-first framework from the LangChain team for building stateful, multi-step agent workflows as graphs.

Good at: Durable, controllable agent flows with explicit state and branching. The right tool once you have outgrown a single prompt loop and need real orchestration you can reason about and test.
The catch: It is a library for programmers, not a no-code builder, so there is a learning curve. It also pulls in the wider LangChain ecosystem, which carries its own weight and churn.

github.com/langchain-ai/langgraph · Pairs with Field Log 006

n8n

A self-hostable workflow automation tool, think of an open Zapier, with strong AI and agent nodes built in.

Good at: Wiring apps, APIs, and language models into automations with a visual editor, self-hosted so your credentials and data stay on your box. The integration library is large and it handles AI steps as first-class nodes.
The catch: It is fair-code, source-available under a Sustainable Use License, not classic open-source. Fine for self-hosting, but read the license before you build a commercial product on it. Big workflows also get hard to maintain.

n8n.io · Pairs with Field Log 006

Edge AI hardware

NVIDIA Jetson Orin Nano

A small developer board with a real GPU for running vision models and smaller language models on-device at the edge.

Good at: On-device inference with CUDA, noticeably more capable than a Raspberry Pi for AI workloads, and a good fit for robotics and camera projects that need to think locally.
The catch: It costs more than a Pi, it runs NVIDIA's own Linux stack (JetPack) that is its own thing to learn, and it draws more power than most people picture when they hear the word edge.

nvidia.com Jetson Orin · Pairs with Field Log 001

Raspberry Pi AI HAT+ (Hailo)

An add-on board that gives a Raspberry Pi 5 a dedicated Hailo neural accelerator for fast local inference.

Good at: Cheap, low-power AI acceleration on hardware you may already own, and a natural fit for local camera and vision work like the kind Frigate does on a Pi.
The catch: It accelerates models compiled through Hailo's toolchain, not anything you please, so you work within its supported-model path. It is an accelerator for specific tasks, not a general-purpose GPU.

raspberrypi.com/products/ai-hat · Pairs with Field Log 004

Coral USB Accelerator

A cheap USB stick with Google's Edge TPU that offloads vision inference from a Pi or a small PC.

Good at: Very low-power, low-cost object-detection acceleration, and the classic partner for a local camera recorder like Frigate running on modest hardware.
The catch: It is TPU-specific. Models have to be compiled for the Edge TPU, so it fits a narrow class of vision models and does nothing for language models. Its tooling and ecosystem have seen limited updates.

coral.ai/products/accelerator · Pairs with Field Log 004

More categories are coming: gadgets, radio and RF gear, and computer hardware. Same rules, same honest catch on every single pick.