LangChain's low-level agent orchestration runtime: durable graphs, streaming, human-in-the-loop, and persistence. About 39,835 GitHub stars.

OpenAI's official Python and TypeScript agent runtime: agents, handoffs, guardrails, sessions, and tracing.

Python serverless compute with a Sandbox API for untrusted agent code, scale-to-zero, and per-second CPU/GPU billing.

Python multi-agent framework with crews, flows, and an Enterprise Cloud from $99/seat. AMP is now that cloud, not a separate product name.

Google's agent SDK in Python, Java, Go, and TypeScript. Docs now live at google.github.io/adk-docs. A2A is a separate protocol repo.

MemGPT's company and OSS agent memory runtime. GitHub letta-ai/letta is ~20.5k stars. Cloud is $20 and $200 per project plus usage.

Ultra-lightweight, self-hosted personal AI agent framework in Python with WebUI, tools, long-term memory, MCP integrations, multi-agent workflows, and scheduled automation.

Anthropic agent library. Python 7,910 stars, MIT.

PrimeIntellect's self-improving coding agent built on a Recursive Language Model (RLM). A persistent IPython kernel as its single tool, a /refine command that lets it improve its own prompts, skills, and memory.

MiroFish is an open-source swarm intelligence engine that spins up agent personas around a news event to surface second-order effects, not a single forecast.

Open-source Python library that makes websites accessible to AI agents, letting them drive a real browser to click, type, and fill forms.

NVIDIA's open-source agent framework that turns an agent into one Python class: methods are actions, docstrings are prompts, at 82.2% SWE-bench Verified.

Praxist is a Fair Source research runtime that runs parallel peers, durable evidence, and Codex or Claude Code skills on a measurable project.

AWS's batteries-included agent harness: one import gives you a tuned general-purpose agent that the team says runs 28 percent cheaper than Claude Code.

An efficient local LLM inference library built around the EXL3 quantization format, running large open models on consumer GPUs with tensor and expert parallelism, and powering the TabbyAPI server.

MIT local gateway that points Claude Code, Codex, Pi, OpenCode, or Cline at 48 providers.

NVIDIA's open-source LLM vulnerability scanner with dozens of plugins and thousands of prompts for probing jailbreaks, prompt injection, and data leakage.

NVIDIA's open-source toolkit for adding programmable guardrails to LLM conversational applications, with Colang flows for input, dialog, retrieval, and output rails.

AI Data Extractor is an MIT Python toolkit that pulls local chat history from Claude Code, Cursor, Codex, and other coding agents into JSONL.

HKUDS open-source registry that turns GUI apps, APIs, and SaaS tools into agent-native CLIs for OpenClaw, Claude Code, Codex, and Cursor.

Microsoft's open-source polyglot agent that researches a repo, then plans, writes, and verifies unit tests: 92.1% completion versus 78.9% for stock Copilot.

Mistral terminal agent. CLI $0 plus Mistral API key. --agent auto-approve / --yolo. Not a $20 seat.

Open-source AI penetration testing tool. Autonomous AI agents that run your app, find vulnerabilities, and validate them with working PoCs. Apache-2.0.

MCP for Blender connects Blender to any LLM over the Model Context Protocol, so an agent can model, texture, and render 3D scenes from a prompt.