AI-Agent - OpenSource is Better

I have been using AI coding assistants for a while now. They all do roughly the same thing: help you write and understand code faster. But the one I keep coming back to is OpenCode. It fits like a glove.

Built for the terminal

OpenCode follows what good CLI tools do: lightweight, minimal interface, but gives you exactly what you need to see what is going on. The output is colour-coded, well laid out, and easy to read at a glance. When it explains what it is doing, I can see the plan clearly before it executes. When it writes code, I can review the diff before applying it.

OpenCode CLI interface showing structured AI response about WordPress wp-login.php

Desktop-based applications often either focus on the wrong things, or there is too much going on to keep the important functionalities in mind. OpenCode feels like a tool designed for people who live in the terminal.

Open.. everything

The OpenCode CLI is open source but also open, full stop. It does not try to lock you into anything. It tries to support as many different platforms, subscriptions etc. I want the flexibility to switch AI provider when there is a deal, a new model or to walk away from something I no longer like. Open source also means the community can contribute, audit, and improve it.

The OpenCode Go subscription

This is not needed to use OpenCode if you already have a subscription. If you don’t and/or you’ve been sitting on the fence for quite some time, maybe waiting for the perfect entry point into AI, I suggest you give it a try.

OpenCode Go costs then $10 per month ($5 for the first month). For this $10 you get $60 worth of usage. This is because usage is measured in dollar value rather than raw request counts.

The limits run in rolling windows to avoid overloading the platform. You get a 5-hour rolling limit of $12, a weekly limit of $30, and a monthly limit of $60. The actual number of requests you can make depends on which model you use — cheaper models get you many more requests, advanced ones fewer.

For example, with DeepSeek V4 Flash I get over 30,000 requests per 5-hour window. That is plenty for daily use. More advanced models like GLM cost more per request, so the same window buys fewer of them — I save those for when I need stronger reasoning.

If I need more usage, I can top up credit or let it fall back to my Zen balance. And once I hit the paid limits, the free models included with OpenCode still work.

There is also OpenCode Zen if you want the flexibility of a wider range of models and providers. But for most of what I do, Go covers it.

Models I actually use

I have settled on a few that cover most of my needs:

DeepSeek V4 Flash — my everyday driver. Fast, great value on credits, good for probably 80% of my interactions.

MiMo V2.5 — when I need vision capabilities. Looking at screenshots, debugging UI issues. Similar value to DeepSeek Flash.

GLM 5.x — for infrastructure work and complex multi-step reasoning. Excellent at planning, but I use it selectively because it consumes credits faster.

The ability to pick the right model for the task rather than being locked into one is exactly what I want. Some providers try to keep you inside their ecosystem. OpenCode lets you choose.

Actively maintained

The tool is actively developed. New models are added regularly, the CLI keeps improving, and the community around it is engaged.

Give it a try

If you live in the terminal and want an AI coding assistant that respects that, OpenCode is worth a try. It has become an indispensable part of my daily workflow.

If you want to give it a shot, you can sign up at opencode.ai (full disclosure: that is my referral link, which will give BOTH OF US $5 of extra credits).

Andrea

Paperclip dashboard showing agent orchestration UI

Paperclip recently caught my attention as it allows you to orchestrate teams of specialised AI agents through a beautiful self-hosted web UI, so I had to give it a try.

What Works Well

The concept is solid: specialized agents for different tasks, all tracked through a ticketing system. The UI looks clean and modern. Navigation feels intuitive. Watching my “CEO” agent hire a UI/UX designer, then a CTO, all coordinating through the interface, was entertaining.

For users who want AI agent orchestration without building everything from scratch, Paperclip provides a polished starting point.

Where It Fell Short for Me

I deployed Paperclip using Docker Compose (my preferred approach for managing services). The setup required some debugging—it always does, of course. Permissions, configs of some utilities not where I was expecting to find them. Probably my fault for trying to optimise too early but, really, I was trying to learn. Nothing insurmountable, but it added complexity not needed at the beginning of a project.

More importantly, I noticed my coding credits depleting steadily while Paperclip was running… with nothing actually happening. My team consisted of a CEO and a designer, happily chatting away, checking tickets in and out, burning credits through heartbeat checks and status polls. The UI showed “agents working” but no real output was being produced.

The pretty interface, it turns out, hides the mechanics. For someone who lives in logs and CLI output, this opacity was a little confusing.

The Simpler Alternative

I stepped back and asked: what am I actually trying to achieve here? Agent orchestration. I then investigated how OpenClaw could solve the same problem and realised I could probably achieve most of what I wanted without the extra software and overhead. Like I said, the UI/dashboard is nice but I can’t just sit there and watch the credits slowly disappear.

If you’re already running OpenClaw, n8n, or similar tools? You’re probably already 80% there with infrastructure you understand end-to-end.

What Paperclip Does Differently

Here’s where Paperclip differs from other harnesses. It’s an orchestrator of harnesses, therefore it can orchestrate multiple different AI harnesses as a single team. As Paperclip puts it “if OpenClaw is an employee, Paperclip is the company”. The employees don’t need to be all of the same type, you could have Claude Code, an OpenClaw and an Hermes Agent all working as part of the same team/organisation.

Paperclip on GitHub — check out the project if you’re curious.

All working together, checking out tickets, assigning work to each other. Paperclip lets you mix and match.

For me, that wasn’t the missing piece. But if you’re building a multi-agent system where different agents need different strengths, it’s worth considering.

If You Want to Try It

The concept shows promise. The project has real potential. If you’re curious about AI agent orchestration and want something with a polished UI, give it a spin.

Learn from my mistakes:

When deploying with Docker, pin a stable release, not latest
Watch your credit usage closely until you understand what it’s doing (maybe set some alerts if you can)
Start with a small project to see results quickly
Understand that there is some complexity going on which is nicely hidden under the UI

If you, on the other hand, are already using some alternative AI harness and all you want is an easier way to manage an AI team/agents, look more deeply into what your software of choice already supports.

Hope it helps!