Collection
Software built around language models rather than bolting them on.
AI coding agent for terminal, editor, and desktop
OpenCode is a coding agent that runs where you already work rather than in a separate chat window. The same session can be driven from a terminal, an editor extension, or a desktop client, and it keeps context across all three.
Turn any website into clean, model-ready data
Firecrawl handles the unglamorous part of feeding the web to a language model: rendering JavaScript, following pagination, stripping navigation and cookie banners, and returning markdown or structured JSON that a model can actually use.
Run open language models locally with one command
Ollama reduced local model hosting from an afternoon of CUDA troubleshooting to a single pull command. It handles model download, quantisation selection, GPU offloading, and memory management, and exposes an OpenAI-compatible endpoint on localhost.
Build and deploy AI agents and RAG pipelines visually
Dify is where a prompt experiment becomes something a team can operate. Workflows are assembled on a canvas, datasets are indexed and versioned, and every deployment gets an API endpoint with logging and cost tracking attached.
Drag-and-drop builder for agents and RAG pipelines
Langflow turns pipeline construction into a graph you can see. Components are Python underneath, so a flow that outgrows the canvas can be exported and run as code rather than rebuilt.
One chat interface for every model you run or connect
Open WebUI is the front end that turns a local inference server into something a household or a team can actually use. It handles accounts, conversation history, document retrieval, and model switching, and speaks to both local runtimes and hosted APIs.
Inference engine that runs LLMs on almost any hardware
llama.cpp is the layer most local model tooling is built on. It is a C and C++ inference engine with no Python runtime, aggressive quantisation support, and hand-tuned kernels for a remarkably wide range of hardware.
OpenAI's assistant and the interface that set the pattern
ChatGPT is the product that made conversational assistants ordinary. Beyond the chat box it now covers file analysis, image generation, web browsing, voice, and custom assistants sharing one account.
Let AI agents drive a real browser
Browser Use gives an agent a real browser rather than an HTTP client. It handles the translation between a model's intent and DOM actions, keeps a structured representation of the page in context, and recovers from the usual navigation failures.
AI completion and chat inside the editor
Copilot put model assisted completion in front of a very large number of developers first, and inline suggestion in the editor remains the interaction most people mean by AI coding.
Autonomous agents that complete real engineering tasks
OpenHands runs agents that plan and execute multi-step engineering work: reading a repository, editing files, running tests, and iterating until a task is done. Each agent works inside a sandboxed runtime rather than on the host machine.
Polished chat interface for many model providers
LobeChat is the most designed of the open chat front ends. Plugins, an assistant marketplace, voice input and output, and vision support are all present, and the interface holds up next to commercial products.
Desktop application for running language models locally
LM Studio made local models approachable for people who do not want a terminal. Browse a catalogue, download a quantisation that fits the machine, and chat, with hardware requirements explained in plain language.
AI dictation that cleans up speech as you talk
Wispr Flow is dictation that accounts for how people actually speak. False starts, filler, and self-correction are removed, and the output is punctuated text rather than a transcript.
Hosted headless browser infrastructure for agents
Browserbase runs the browsers so you do not have to. Sessions, proxies, captcha handling, and stealth configuration are managed, which removes the least interesting and most fragile part of building anything that scrapes or automates the web.
Speech recognition that runs anywhere, with no runtime
whisper.cpp is a C and C++ port of the Whisper speech recognition model with no Python dependency, quantised weights, and hardware acceleration across Apple Silicon, CUDA, and plain CPU.
An editor rebuilt around model assisted editing
Cursor forked VS Code rather than extending it, which let it change the editing model itself. Multi-file edits, codebase-wide context, and a tab completion that predicts the next edit rather than the next token are the result.
Anthropic's assistant, strongest on long documents and code
Claude is positioned around careful reasoning over long inputs. The large context window makes whole-repository and long-document work practical in a way that shorter context assistants cannot match.
A control layer for customer-facing AI agents
Parlant addresses the problem that stops most support agents from shipping: a model that is helpful ninety-five percent of the time is not safe to put in front of customers. It manages behaviour through explicit, context-matched guidelines rather than one long prompt.
Local voice dictation with AI cleanup for macOS
VoiceInk transcribes speech on the machine itself using a local Whisper model, then runs the raw transcript through a cleanup pass that removes filler and applies punctuation.