Made a massive curated list of 260+ AI agents & tools — heavy focus on open-source, self-hosted, and local-first options

Posted by Caramaschi@reddit | LocalLLaMA | View on Reddit | 8 comments

I put together what I think is the most comprehensive list of AI agents and frameworks available right now, with a big emphasis on open-source and self-hosted tools.

https://github.com/caramaschiHG/awesome-ai-agents-2026

Some highlights for this community:

**Local LLM Runners:** Ollama (162k stars), llama.cpp, vLLM, LM Studio, Jan, LocalAI, GPT4All, Llamafile

**Self-hosted agents:** OpenClaw (the 9k→188k stars phenomenon), Open WebUI, LibreChat, LobeChat, Anything LLM, DB-GPT

**Open-source frameworks:** Smolagents (HuggingFace), DeerFlow (ByteDance, #1 trending), LangGraph, CrewAI, AutoGen, Mastra

**Open-weight models for agents:** Llama 4, Qwen 3 (MCP-native!), DeepSeek V3/R1, GLM-4 (lowest hallucination), Gemma 3, Phi-4

**Open-source video gen:** Wan 2.1 (self-hostable, no limits), HunyuanVideo, LTX Video

**OSS voice:** LiveKit Agents, Rasa, Pipecat, Vocode

**Browser infra:** Browser Use (what Manus uses under the hood), Skyvern, Agent S2

Plus vector DBs (Chroma, Qdrant, Milvus, Weaviate), RAG engines (RAGFlow, Pathway), safety tools (NeMo Guardrails, LLM Guard), and a lot more.

CC0 licensed. PRs welcome. What am I missing?

[-]

Conscious_Chapter_93@reddit

Great list. One category I'd add is agent operations/control plane.

Once you start running more than one agent locally, the pain shifts from "which framework?" to "what is installed, what is running, what tools can it call, what changed, and how do I stop it safely?"

I'm building Armorer for that layer: https://github.com/ArmorerLabs/Armorer

Great list. One gap in the agent infrastructure section — payment rails for agents. If your agent needs real-time crypto data (price, whale activity, gas, TVL etc), AgentPay lets it pay per call in USDC via x402 instead of holding API keys. No subscription, budget-capped. Works as an MCP server too: https://github.com/romudille-bit/agentpay

niga_chan@reddit

this is actually a really solid list, nice work putting this together

one thing that feels slightly under-discussed is what happens when you actually try to run these agents at scale locally

we’ve been experimenting on that side and were able to push \~4.5k standby agents on a single node, and pretty quickly the bottleneck shifts from models/tools → orchestration + memory per agent

feels like there’s a missing layer around “agent infrastructure” that sits between all these frameworks and real-world usage

curious if you’ve come across anything focused more on that layer

APIS_AI@reddit

We built ContextGate: a dynamic HUD-style context header for agents.

Trusted local state up top, untrusted remote data kept separate, replaced each turn instead of bloating transcript history.

https://github.com/APIS-AI/contextgate

A trusted DESKTOP header lets an agent keep live notes and runtime state without re-appending them through chat history.

That means more token budget stays available for the actual coding task.

My GPT-5.4 coding agent has been running for days and is still at 3% context.