AI tools, agents, models, and research catalog

What it is: vLLM remains a high-throughput inference server with fast-growing support for new architectures and deployment patterns.

Why it matters: Use this if you run LLM APIs today and care about throughput, latency, and memory efficiency.

What it is: OpenAI's Codex is an official, open-source terminal coding agent that turns prompts into concrete code operations (open/edit/run/test), integrates approval-aware execution, and supports multiple AI providers via OpenAI-compatible APIs.

Why it matters: Useful for solo developers and teams that want a local terminal coding assistant with a narrow execution scope, explicit approvals, and straightforward model routing before turning on heavier agent frameworks.

What it is: AG2 is an active open-source AgentOS with protocol-level tools, multi-agent orchestration, and explicit A2A interoperability design.

Why it matters: Use it when you need coordinated AI agents with stronger conversation-state handling than ad-hoc chains.

What it is: anomalyco/opencode is an MIT-licensed open-source coding agent that ships as a local CLI with a hosted UI, broad language docs, and a real install path instead of another prompt pack or thin wrapper around an existing editor assistant.

Why it matters: Useful for developers who want a serious open coding-agent stack they can run, inspect, and adapt without locking the workflow inside one proprietary IDE.

What it is: AgentScope is a production-oriented Python framework for multi-agent systems with explicit tooling, state handling, and MCP-compatible execution paths.

Why it matters: Use AgentScope when you need structured handoffs, reusable agent primitives, and observable orchestration instead of ad-hoc scripts.

What it is: SGLang offers a fast-serving compiler/runtime path for LLM and multimodal workloads with active release velocity.

Why it matters: Useful for teams needing a flexible serving stack that supports advanced routing and model-serving optimization.

What it is: Gemini CLI is Google's Apache-2.0 terminal agent that gives developers direct Gemini access with a free tier, 1M-token context, built-in shell and file tools, web fetching, and MCP extensibility.

Why it matters: Useful for developers who want a first-party terminal agent from a major model vendor, especially if they want generous free usage and a fast way to test Gemini on real repo and command-line tasks.

What it is: Open WebUI is a source-available self-hosted AI workspace that runs offline, connects to Ollama and OpenAI-compatible APIs, and packages chat, voice and video, agents, and model management into one interface.

Why it matters: Useful for teams and power users who want a polished front end for local or mixed-model AI stacks without handing the whole workflow to a closed hosted dashboard.

What it is: ruflo is a high-adoption agentic meta-harness for orchestrating multiple AI agents with native Claude Code / Codex integrations and memory workflows.

Why it matters: Useful for teams shipping practical multi-agent systems who need swarming, memory, and tool handoff without building everything from scratch.

What it is: OpenAI's multilingual large-v3 Whisper checkpoint for speech recognition, language identification, and speech translation.

Why it matters: Whisper large-v3 is a mature baseline with extensive ecosystem support and a clear official model card, making it useful well beyond its original release cycle.

What it is: Apache-2.0 canonical Agent Client Protocol (ACP) -- open standard (current stable protocol version 1) for connecting any code editor to any coding agent via JSON-RPC; ships first-party SDKs in 5 languages

Why it matters: Most coding-agent developers today who want their agent to plug into a code editor (Zed / VS Code / JetBrains) have been either (a) writing custom JSON-RPC bridges per editor (no canonical protocol), (b) reaching for non-standard integrations like stdin / stdout file-watch that require custom agent-side code, or (c) maintaining separate transport layers per editor.

What it is: awslabs/mcp is a maintained catalog of production AWS MCP servers with documented setup paths for retrieving and acting on AWS context safely from agent clients.

Why it matters: Useful for teams on AWS who need practical MCP integrations without hand-building service-specific glue.

What it is: BerriAI/litellm is a production-ready proxy and toolkit that normalizes calls across 100+ LLM providers while adding observability and guardrails.

Why it matters: Useful if you need a single, auditable routing layer instead of wiring each model provider separately.

What it is: llama.cpp is an MIT-licensed C/C++ runtime for local LLM inference with GGUF models, server builds, Docker packaging, and a very large maintainer/user footprint. Current GitHub metadata shows the canonical repo under ggml-org, active maintenance, and a raw MIT license.

Why it matters: Local inference is still one of the most practical ways to test open models privately and cheaply. llama.cpp matters because it is the common runtime layer behind many desktop, server, and edge experiments instead of a narrow demo wrapper.

What it is: GitHub's official MCP server exposes GitHub workflows, repos, issues, and code context through a standards-based tool surface for AI assistants, so tooling can be connected via a single Model Context Protocol endpoint.

Why it matters: Useful for teams already living in GitHub who need stable MCP access to repository data and issue/workflow actions without a custom server layer.

What it is: Kilo Code is an MIT-licensed open-source coding agent that runs in VS Code, JetBrains, and the CLI, lets users pick from 500+ models with provider pricing, and supports switching models mid-task.

Why it matters: Useful for developers who want one serious open coding-agent surface across editor and terminal workflows instead of locking themselves into a single hosted assistant experience.

What it is: AnythingLLM is a MIT-licensed local-first workspace for document chat, agent workflows, multimodal work, and self-hosted team use, with active releases, broad model support, and a much lower setup burden than piecing together a private RAG stack by hand.

Why it matters: That matters because a lot of people want one place to run local or self-hosted AI against their own files without stitching together a vector store, chat UI, agent layer, model router, and access controls from scratch.

What it is: Rapid-MLX is an Apple Silicon local inference engine that serves open models through an OpenAI-compatible API, adds native tool calling and prompt caching, and plugs directly into Cursor, Claude Code, Aider, and other coding-agent clients.

Why it matters: Useful for Mac users who want a faster local-model path than the usual Ollama setup and a cleaner drop-in endpoint for coding agents, automation, and private day-to-day AI work.

What it is: Sandbox Agent is a server that runs inside an isolated environment and exposes coding agents such as Claude Code, Codex, OpenCode, Cursor, Amp, and Pi over HTTP or SSE instead of forcing every team to hand-roll remote TTY control and session management.

Why it matters: Useful for teams that want coding agents to run in real sandboxes rather than on a developer laptop. It makes the remote-control layer reusable, which is a much more practical step toward safe agent execution than another local-only wrapper.

What it is: Stability AI's established SDXL base checkpoint for high-resolution text-to-image generation and downstream fine-tuning.

Why it matters: SDXL remains a practical ecosystem baseline with extensive tool support, although its OpenRAIL terms and synthetic-image risks still need to be read.

What it is: bitsandbytes remains one of the highest-signal local quantization libraries for running larger LLMs under tighter memory budgets on single-node GPUs.

Why it matters: Useful for practical model deployment when GPU memory is the limiting factor and you need production-ready quantization patterns.

What it is: Cherry Studio is a cross-platform desktop AI client that unifies frontier and local model providers, bundled assistants, document chat, and MCP support in one ready-to-use app.

Why it matters: Useful for power users who want one polished desktop surface for Claude, Gemini, OpenAI, Ollama, LM Studio, and document-heavy workflows instead of juggling separate apps and browser tabs.

What it is: Egonex-AI/Understand-Anything is an MIT-licensed code and knowledge-graph workspace that turns repositories, docs, and notes into an explorable graph with search, chat, and agent-friendly context for Claude Code, Codex, Cursor, Copilot, and similar tools.

Why it matters: Useful for developers who keep losing time to repo orientation and context reconstruction across large codebases, docs, and internal notes.

What it is: RAGFlow continues as a mature open-source engine for context retrieval plus agent workflows, with recent updates for model-provider flexibility and model configuration control.

Why it matters: Adopt RAGFlow when you already run RAG-heavy products and want a stronger orchestration layer for connectors, retrieval, and agent context handling.

What it is: Kreuzberg is a Rust-core document intelligence framework that extracts text, metadata, tables, OCR results, and code structure from 96 file formats and 306 programming languages, with bindings for major runtimes plus CLI, REST, and MCP surfaces.

Why it matters: Useful for teams building document-heavy AI workflows that need one serious extraction layer instead of a pile of single-format parsers and ad hoc OCR scripts.

What it is: LMCache focuses on KV-cache acceleration with AMD/CUDA support and a clear optimization-oriented release cadence.

Why it matters: It targets the same production pain point as most teams: rising GPU cost from context-heavy workloads.

What it is: MCP Gateway is Microsoft's reverse proxy and management layer for MCP servers, adding session-aware routing, auth controls, server lifecycle management, and Kubernetes-friendly deployment paths for teams that need more than a single local MCP process.

Why it matters: Useful for platform teams that are moving from one-off MCP experiments to shared infrastructure: it gives them a cleaner way to route tool traffic, keep sessions sticky, and manage multiple servers without inventing the control plane from scratch.

What it is: TokenTracker is a local-first usage monitor that auto-collects token counts and spend across 22 AI coding tools, then surfaces the totals in a dashboard plus native desktop tray and widget views.

Why it matters: Useful for anyone juggling Claude Code, Codex, Cursor, Gemini CLI, Roo Code, Goose, and similar tools: it turns scattered usage logs into one readable local view so teams can spot cost drift before it becomes a budgeting problem.

What it is: nexu-io/open-design is an Apache-2.0 local-first native desktop workspace that turns the closed 'Claude Design' style loop into an open toolchain: design systems, skills, plugins, sandboxed prototype previews, and export paths for HTML, PDF, PPTX, MP4, and more from one editable project surface.

Why it matters: Useful for designers, founders, and creative builders who want agent-assisted prototypes and presentation assets without being locked into a hosted black box or a browser-only workflow.

What it is: Off Grid is an MIT-licensed offline AI app for iOS, Android, and macOS that runs chat, vision, image generation, voice, and retrieval on-device instead of sending sessions to a cloud model.

Why it matters: Useful for privacy-conscious people who want a real offline AI app they can carry on a phone or laptop instead of a demo that still depends on a remote model API.