how is this even allowed i dropped my $200/mo claude plan for deepseek harness, it’s basically free the whole thing runs on one simple rule: everything is a plugin [setup takes 3 minutes, here’s the playbook] 1. clone/install the repo called deepseek-harness 2. skim the architecture docs 3. that’s... it, you’re live bookmark this and hand it to your agent right now.show more

How does it compare to something like Hermes? I use Hermes now for basic things, setting up VPS, self hosting, learning JavaScript, etc. I don't have the crazy use cases a lot of you guys do.

Love this! The plugin‑first design is really interesting. Curious how you see it comparing or even coexisting with Hermes Agent. 👀 A harness like this could slot in as a specialized coding layer.

Do you have a link to the harness?

@grok this looks familiar... Magi system anyone? #evangelion

@cyrilXBT wild how everyone’s finding these crazy cost-saving hacks. got me thinking if i'm overpaying for stuff too 🤔

The math only works because of DeepSeek's $0.07 per million token prompt caching rate. Without an 80% plus KV-cache hit rate, an agent loop running 50 tool invocations per task burns through token budgets faster than a flat $200 Claude tier. The hidden bottleneck in an everything-is-a-plugin setup is schema ordering. If your harness appends plugin schemas dynamically, it mutates the system prompt prefix and breaks cache alignment on every call.

Bookmarked Thanks champ

'everything is a plugin' sounds clean until one plugin swallows an exception and the next step silently no-ops. took me 3 hours to realize my pipeline wasn't failing, it was succeeding with empty output.

This is an amazing harness

"everything is a plugin" is the scary part, not the free part. If adapters and tools are swappable, lock-in moves from weights to distribution. Whoever ships the default agent harness for VS Code / CLI wins the next layer. OSS is how you take it from the frontier labs.

cheap agent stacks are cool, just audit those plugins carefully

Deep seek is really doing great it's cheap and also give output like top model

Moving from $200/mo closed-source wrappers to open harness setups is becoming the standard pipeline migration. You’re not just saving costs on inference; you get total control over tool-calling execution and context routing.

smart move, this could save a ton of money

Looked at the docs last week. Plugin-first design is sound. Still keeping Claude for complex reasoning though, that gap hasn't closed yet.

Free open harness replacing a $200/mo plan in three minutes.The agent layer is commoditizing fast.

Plugin architecture is the durable bet: model, tools, loops, sandboxes, and sessions can evolve independently. Long-running failure recovery will decide whether the open harness stays pleasant to extend.

$200 saved

you'll like gitlawb's openclaude

Hey Everyone. Who's passionate about building something and involved in deep thesis work but in short of AI tools and resources? I got you! I found a website that gives 175$ claude opus-4.8 and opus-5, gpt-5.6 for free and includes daily-bonus of 25$ link:

thanks for sharing this 💯

Shows how scam Claude is

Is it sending data anywhere weird

did it actually replace claude for agent work, or just the easy plugin calls?

Basically free, that's wild

just cloned it and the whole thing was live in under 3