Open source agent stack that actually works in 2026 (no hype)

Posted by virtualunc@reddit | LocalLLaMA | View on Reddit | 24 comments

been running this setup for a few months and wanted to share what actually works vs whats just github stars

hermes agent (24k stars, MIT) - runs on your own machine or vps, connects to telegram/discord/whatsapp, persistent memory stored locally in sqlite. your data doesnt leave your network. pair it with ollama and local models for fully offline operation. the memory layer is what separates it from everything else.. your agent on day 30 actually knows your projects and preferences

everything claude code (128k stars) - 27 agents, 64 skills, 33 commands. installs directly into claude code or hermes as a skill set. covers planning, code review, security audits, TDD, token optimization

dify (130k stars) - self hostable visual workflow builder. raised $30M. production ready, not a toy. runs on your own hardware via docker

onyx (17k stars) - self hosted private search across your google drive, notion, confluence, slack. nothing leaves your servers

repomix (21k stars) - one command to pack your entire codebase into a single file claude can read. solves the context problem instantly

flowise (30k stars) - lighter weight alternative to dify if you want drag and drop agent building without the complexity

the combo that works best for me: hermes for persistent memory and 24/7 task handling, repomix for giving it full codebase context, everything claude code for the skill library

all MIT or Apache licensed, all self hostable, all actively maintained

[-]

Enough_Big4191@reddit

this kind of stack works until u have to debug why the agent made a bad call, then the seams between tools start to hurt. i’ve had cases where memory said one thing, enrichment another, and the agent just picked one confidently. I’d sanity check it by logging decisions end to end for a week and see where identity or context drifts, that’s usually where these setups break first.

thetomsays@reddit

Any recommendation for either a tool/service with UI or a traditional log with LLM investigator?

justbeast@reddit

I suspect you wanted to post this in reply to a different comment..

No, I meant to ask about systems/tools for the sanity check and logging process.

What does enrichment meant, in this context?

HackerSpear@reddit

The problem that I have had with repomix that it creates too detailed resume of my projects, and gets close to 1B tokens file that none but Gemini could read.

AHHHH_AHHHHHHHH@reddit

This the type of content this sub needs. Thanks homie

Professionalist1501@reddit

hey man i've a work for you and i'm not a scammer so please consider replying i can show you proofs

MrPanache52@reddit

Scammer

Distinct-Opposite779@reddit

Good list. One thing worth adding on the memory side, storing conversations is the easy part. The gap I've hit is retrieval quality. When your agent has hundreds of past interactions, vector similarity alone starts returning stuff that's semantically close but contextually wrong. Like the agent remembers you mentioned a deadline three months ago and treats it as current. Or it pulls a preference you explicitly changed. The structured knowledge extraction layer matters more than the storage layer once you're past the demo phase. Hermes local SQLite is fine for getting started but keep an eye on how retrieval holds up at scale.

Are there any solutions that you've found to these storage challenges?