Build stateful, multi-step agent workflows with explicit graph control.
Why notableThe graph model beats linear chains for agents that need to loop, branch, and retry. Active development, strong production adoption.
01
Agent Frameworks
Autonomous and multi-agent orchestration. The fastest-moving category in Q2 2026.
Multi-agent conversation framework. Agents talk to each other to complete tasks.
Why notablev0.4 shipped a complete rewrite with async-first architecture. Still the most cited framework in multi-agent research.
Role-based agent orchestration — assign agents jobs, watch them collaborate.
Why notableFastest-growing framework for non-research use cases. Simple API, large community, integrates with most LLM providers out of the box.
Minimal agent library built around code execution as the default action.
Why notableUnder 1,000 lines of core code. Code-first agents outperform tool-call agents on most benchmarks.
OpenAI's own lightweight multi-agent SDK. Replaces Swarm.
Why notableOfficial library, active maintenance, clean handoff primitives. If you're building on the OpenAI API, this is now the canonical starting point.
Build, run, and manage agent platforms. Focused on production deployment.
Why notableTargets the gap between "runs in a notebook" and "runs in production at scale." Good observability hooks built in.
Agent framework built on top of Pydantic. Type-safe LLM outputs by default.
Why notablePydantic already owns structured output parsing. This extends that to full agent workflows. Production-grade from day one.
Design principles for production-grade LLM-powered software. Not a framework, a spec.
Why notableThe clearest articulation of what separates toy agents from production ones. Essential reading before picking a framework.
Multi-agent framework that simulates a software company — PM, engineer, QA all as agents.
Why notableResearch-forward, but increasingly practical. Best for complex code generation tasks that require multiple review passes.
Frontend stack for embedding agents and generative UI in React apps.
Why notableMakers of the AG-UI Protocol. The only serious framework for building agent interactions directly into web UIs.
02
LLM Serving and Inference
Run models faster and cheaper. The infrastructure layer for everyone building on open weights.
Get open-weight models running locally in one command.
Why notableSupports Kimi-K2.5, GLM-5, MiniMax, DeepSeek, Qwen, Gemma, and more. The easiest on-ramp to local inference by a wide margin.
LLM inference in pure C/C++. Runs on CPUs, Apple Silicon, and consumer GPUs.
Why notableThe engine behind most local inference tools. Quantization quality and speed continue to improve quarter over quarter.
High-throughput, memory-efficient LLM serving with PagedAttention.
Why notableThe default choice for serving open-weight models at scale. Continuous batching, OpenAI-compatible API, multi-GPU support.
High-performance serving for LLMs and multimodal models.
Why notableBenchmarks consistently faster than vLLM on certain workloads. Worth testing if you are latency-sensitive.
Official inference framework for 1-bit LLMs.
Why notable1-bit quantization is the most aggressive size reduction available. BitNet makes it practical on commodity hardware.
Unified API proxy for 100+ LLM providers in OpenAI-compatible format.
Why notableSingle integration point for every major provider. Cost tracking, load balancing, and guardrails included. Production-proven.
Run LLMs, vision, voice, image, and video models on any hardware without a GPU.
Why notableThe only single-binary solution for running all modalities locally. Privacy-first alternative to cloud APIs.
03
Fine-Tuning and Training
Adapt open-weight models to specific domains and tasks.
Unified fine-tuning interface for 100+ LLMs and vision-language models.
Why notableLoRA, QLoRA, full fine-tune, and RLHF all in one tool. ACL 2024 paper. The most complete fine-tuning harness available.
Fine-tune and run open models with a web UI. Supports Gemma 4, Qwen 3, DeepSeek.
Why notableDramatically reduces memory footprint vs. standard training. Studio UI makes it accessible without writing training scripts.
Parameter-efficient fine-tuning. LoRA, prefix tuning, prompt tuning, and more.
Why notableThe canonical library for LoRA. Integrates with Transformers, Diffusers, and Accelerate. Used by nearly every other fine-tuning tool.
Config-driven fine-tuning for LLMs. Write YAML, not training loops.
Why notableBest option for teams who want reproducible, version-controlled training runs without custom code.
Build fine-tuning datasets from unstructured content with LLM assistance.
Why notableThe dataset problem is often the real bottleneck. This automates the extraction and formatting step.
No-code GUI for fine-tuning LLMs.
Why notableLowers the barrier for non-ML teams. Export to GGUF for local deployment. Solid option for domain-specific model builds.
End-to-end platform for LLM training, fine-tuning, and alignment.
Why notableEnterprise-grade. Used for training production models, not just fine-tuning. Requires NVIDIA hardware but has no ceiling on scale.
PyTorch-native fine-tuning library. No abstractions above the framework.
Why notableMaximum control for teams that need to customize the training loop. Official PyTorch project with strong long-term support.
04
RAG and Vector Infrastructure
Retrieval-augmented generation and the databases that make it work.
The leading document agent and OCR platform for RAG pipelines.
Why notableMoved well beyond naive chunking. Structured retrieval, agent-driven queries, and a large integrations ecosystem.
Universal memory layer for AI agents. Persistent, queryable agent memory.
Why notableSolves the statelessness problem without requiring a custom vector store setup. Integrates with most agent frameworks.
High-performance, cloud-native vector database built for ANN search.
Why notableBattle-tested at scale. Handles billions of vectors. The production-grade option when Chroma becomes a bottleneck.
High-performance vector database and search engine, written in Rust.
Why notableRust performance, strong filtering capabilities, and a clean API. The preferred choice for teams prioritizing speed and reliability.
Search infrastructure for AI. Embeddable, open-source vector store.
Why notableFastest path from zero to working RAG. Runs in-memory or persisted. The default starting point for most RAG prototypes.
Modular pipelines for context-engineered, production-ready LLM applications.
Why notableExplicit control over retrieval routing, memory, and generation. Built for teams that outgrow LangChain's abstractions.
Graph-based RAG. Builds knowledge graphs from documents before retrieval.
Why notableOutperforms standard RAG on complex, multi-hop questions. Slower to index but meaningfully better answers.
Notebook-based showcase of advanced RAG patterns with runnable code.
Why notableNot a library — a learning resource. The fastest way to understand what's beyond naive chunking.
All-in-one AI productivity tool. On-device, privacy-first RAG with no setup.
Why notableThe non-technical user's path to private RAG. Desktop app, no cloud requirement, works with local and remote models.
05
Coding Agents and IDE Assistants
AI systems that write, review, and execute code.
Autonomous coding agent as SDK, IDE extension, or CLI.
Why notableDeep VS Code integration, full file system access, terminal execution. The most capable open-source coding agent in active use.
AI pair programming in your terminal. Works with any Git repo.
Why notablePolyglot, diff-aware, commit-native. Benchmarks consistently high on SWE-bench. The terminal-native alternative to IDE extensions.
Open-source software development agent. Reads, writes, and runs code autonomously.
Why notableFull sandbox environment, browser control, and shell access. The open-source closest approximation to Devin.
Source-controlled AI code checks, enforceable in CI.
Why notableShifted from pure IDE autocomplete to CI-enforced AI review. Adds AI gatekeeping to the PR pipeline.
OpenAI's CLI coding agent. Point it at a task, it writes and runs code.
Why notableOfficial OpenAI release. Sandboxed execution, multi-file edits, works with any codebase.
Open-source AI agent bringing Gemini directly into your terminal.
Why notable104,000 stars in weeks after launch. The fastest-rising coding CLI of Q2 2026. Gemini 2.5 Pro backend with 1M token context.
06
Image Generation
Open-source image synthesis tools for professionals and builders.
The original Stable Diffusion web interface. Still the most widely used.
Why notablePlugin ecosystem is enormous. Every new model, ControlNet, and sampler ships support here first.
Node-based Stable Diffusion UI and inference backend.
Why notablePipelines are graphs — infinitely composable, exportable, and reproducible. Industry standard for production image workflows.
Simplified Stable Diffusion UI focused on prompting and generating.
Why notableZero config, just prompt and generate. Best quality-to-effort ratio for users who don't need node control.
Professional creative engine for Stable Diffusion. Industry-grade web UI.
Why notableThe production tool used by professional artists and studios. Canvas, workflow editor, and team features.
Optimized fork of AUTOMATIC1111 with lower VRAM usage.
Why notableRuns newer models (Flux, SD3.5) faster on consumer hardware. Drop-in replacement for A1111 with meaningfully better performance.
State-of-the-art diffusion model library for image, video, and audio.
Why notableThe code layer that most serious tools build on. Start here to understand how generation works or build custom pipelines.
Diffusion model inference in pure C/C++. Supports Flux, Wan, SD, Qwen Image.
Why notableRuns on CPU and low-end hardware. The llama.cpp equivalent for image generation.
07
Video Generation
Open-source video synthesis and motion models.
Open and advanced large-scale video generative models from Alibaba.
Why notableThe highest-quality open-weight video model available as of Q2 2026. Outperforms most commercial APIs on standard benchmarks.
Real-time-class video generation. Official Lightricks repository.
Why notableFirst open model to hit near-real-time generation speeds. The last_frame_uri parameter enables true seamless loops.
Animate any personalized text-to-image model without specific tuning.
Why notableThe original plug-and-play motion adapter. Still the reference implementation for controlled animation from image inputs.
Open-source video generation models from Tsinghua University.
Why notableCogVideoX series delivers strong text-to-video quality from a research lab without a commercial API paywall.
Open-source attempt to replicate and improve on OpenAI's Sora.
Why notableTraining code included. The only open project that lets you train video generation models at scale.
High-fidelity video generation models from Genmo.
Why notablePrioritizes motion quality and physical plausibility over raw resolution. Good for realistic motion content.
08
Voice and Audio
Open-source TTS, STT, voice conversion, and audio generation.
Robust speech recognition via large-scale weak supervision.
Why notableStill the most accurate open ASR model. Multi-language, multi-task, runs locally. The baseline everything else is compared to.
Whisper re-implementation with CTranslate2. Significantly faster inference.
Why notable4x faster than the original Whisper with comparable accuracy. The production choice for transcription workloads.
Whisper with word-level timestamps and speaker diarization.
Why notableAdds the two features standard Whisper lacks. Indispensable for subtitle generation and meeting transcription.
Flow-matching TTS — ultra-realistic speech synthesis.
Why notableAmong the best open-source TTS models on naturalness benchmarks. Fast inference, clean API.
Multi-lingual large voice generation model from Alibaba.
Why notableSupports zero-shot voice cloning across languages. Full training and deployment stack included.
TTS model capable of generating ultra-realistic dialogue in one pass.
Why notableDialogue-native — handles conversation, interruptions, and speaker transitions that break standard TTS systems.
82M parameter TTS model. Fast, local, surprisingly high quality.
Why notableRuns on CPU in real time. The lightweight TTS option when ElevenLabs latency or cost is a problem.
Framework for voice and multimodal conversational AI.
Why notableHandles the real-time audio pipeline — VAD, STT, LLM, TTS, all stitched together. Building voice agents without pipecat is building your own media server.
Framework for realtime voice AI agents with video and audio.
Why notableLiveKit infrastructure plus agent SDK. The production path for voice agents that need WebRTC-grade reliability.
State-of-the-art open-source TTS with voice cloning.
Why notableOne of the best-sounding open TTS models. Fast inference, multilingual, active development.
09
MLOps and LLM Observability
Track what your models do in production. Catch regressions. Control costs.
Open-source LLM engineering platform. Observability, evals, prompt management.
Why notableThe most complete open-source LLMOps tool. OpenTelemetry-based tracing, session-level analysis, prompt versioning. YC W23.
Open-source AI engineering platform for agents, LLMs, and ML models.
Why notableDebugging, evaluation, monitoring, and cost control in one platform. LLM tracing is now production-grade.
One-line LLM observability. Monitor, evaluate, and experiment.
Why notableLiterally one line of code to instrument. Best time-to-first-insight of any tool in the category. YC W23.
Observability and testing for AI agents. Session replay, cost tracking, evals.
Why notableAgent-specific. Tracks multi-step sessions, not just individual LLM calls. The Datadog equivalent for agent workflows.
Open-source observability for GenAI based on OpenTelemetry.
Why notableStandard-based. If your stack already uses OpenTelemetry, this is the cleanest integration path.
Full-lifecycle AI agent optimization — development, debugging, eval, monitoring.
Why notableOne of the few tools that covers the full loop from prompt to production. Strong debugging UI for complex agent pipelines.
10
Browser and UI Automation
AI agents that operate the web.
Make websites accessible for AI agents. Automate tasks online.
Why notableThe dominant Python library for LLM-driven browser control. 94,000 stars is not inflated — this is the tool the community converged on.
SDK for browser agents. Built for TypeScript.
Why notableThe TypeScript-native alternative to browser-use. Robust session management, cloud-friendly, strong abstraction layer.
Automate browser-based workflows with AI. Vision-first approach.
Why notableUses vision to understand pages instead of DOM parsing. Works on sites that break selector-based automation.
Parse and understand any UI screenshot for agent interaction.
Why notableConverts arbitrary UI screenshots into structured data an agent can act on. The universal UI understanding layer.
Open-source multimodal AI agent stack from Bytedance.
Why notableHandles the full stack from visual understanding to action execution. Production-grade, open-sourced by Bytedance.
Chrome DevTools for coding agents via MCP.
Why notableOfficial Chrome team release. Gives agents direct access to DevTools protocol — debugging, network inspection, DOM access.
Playwright as an MCP server. Browser automation for Claude and other agents.
Why notableOfficial Microsoft release. The standard way to wire Playwright into Claude Code and other MCP-compatible agents.
Natural language interface for computers. Code execution, file system, browser.
Why notable63,000 stars, years of production use. The original "LLM with computer access" project. Still the most general-purpose option.
11
MCP Servers
The Model Context Protocol ecosystem — servers that give agents real-world capabilities.
The official MCP server reference collection from Anthropic.
Why notableFilesystem, GitHub, Slack, memory, fetch, and more — all in one repo. The canonical starting point for understanding what MCP can do.
GitHub's official MCP server. Full GitHub API access for agents.
Why notableOfficial GitHub release. Agents can read issues, create PRs, search code, and manage repos without custom integration work.
Up-to-date code documentation for LLMs and AI code editors.
Why notableSolves the hallucination-from-stale-docs problem. Pulls live, version-specific documentation into the agent's context at query time.
MCP toolkit for coding — semantic retrieval and editing capabilities.
Why notableSemantic code search and structured editing over an MCP interface. The IDE capability layer for coding agents.
Autonomous agent for deep research on any topic using any LLM.
Why notableMCP-compatible research agent. Gathers, synthesizes, and reports from the web autonomously.
AI agents, MCPs, and workflow automation with ~400 MCP server integrations.
Why notablen8n-style workflow automation with deep MCP integration. The broadest set of ready-made agent action connectors.
MCP server for building n8n workflows via Claude Code or Cursor.
Why notableLets agents construct n8n automation flows through natural language. Practical bridge between agent instructions and workflow automation.
Save this page
Star repos in batches by category. The star is the bookmark. When you're ready to solve a problem in a given category, you'll have a pre-filtered shortlist instead of starting from trending. Revisit at the start of Q3 2026 — star counts shift fast, and the repos that matter will have committed steadily through the quarter.
Check the license before you build a product on any repo. MIT and Apache 2.0 are generally safe for commercial use. AGPL-3.0 has copyleft requirements that matter at the product layer. CC BY-NC is non-commercial only.
Part of the AIgent Library series