OpenCode Shenanigans EP1: My Hacks to Turbocharge Dev Life!

Image from https://opencode.ai/

**OpenCode** is an open-source AI coding agent. It’s available as a terminal-based interface, desktop app, or IDE extension.

Perfect for devs who want to keep things private and flexible without Big Tech breathing down their neck.

OpenCode vs Claude

OpenCode is a rebel: it runs locally, executes code in multiple languages (primarily Python but extensible via system commands), supports over 100 LLMs through LiteLLM — including local ones via Ollama — for full customization and no API costs. Claude emphasizes safety and structure, while OpenCode delivers raw, unfiltered power (depending on your chosen model) for advanced scripting and automation.

In regions like Hong Kong, where direct access to **claude.ai is blocked (requiring a VPN for the web interface), OpenCode shines with fully local or proxy-based models — no regional restrictions. While workarounds exist for Claude (e.g., VPNs or third-party platforms), OpenCode** offers a smoother overall experience: easy model swapping, a clean interface, and complete freedom from geographic or API drama.

Trial Steps to OpenCode

Flipping OpenCode into myultimate dev wingman — smooth automation with brains, dodging pitfalls like model ghosting and limit whines.

Plugins

oh-my-opencode is like the oh-my-zsh of AI terminals — your ultimate hype man for agents that orchestrate tasks, snoop codebases, and delegate like a pro.

Its killer move? Streamlining wild workflows, like plotting code overhauls or googling external gems, via ready-to-roll Model Context Protocols (MCPs) such as Context7, Exa Code Search, and Web Search. It auto-picks agents and juggles parallel madness, slashing lag for those “just one more iteration” dev binges. Basically, it turns OpenCode into a speed demon.

opencode-supermemory? It’s OpenCode’s elephant brain upgrade — keeps context alive across sessions like a clingy ex. Stashes project deets (tech stack, architecture quirks) and your quirky prefs in neat scopes, dishing out smarter replies without you repeating yourself like a broken record.

Gold for marathon projects: Fire up supermemory-init, let it sleuth and hoard insights, and boom — no more “wait, what was that again?”

Agents

To pimp out oh-my-opencode, I cooked up Hermes — a zippy custom agent that’s basically a traffic cop for tasks. It sorts queries and hands ’em off smartly, skipping pointless parallel party crashes for those tiny “small small” gigs like fast queries or minor tweaks.

Result? oh-my-opencode gets a responsiveness boost, ditching the fluff for everyday dev wins.

Hermes: Because who needs drama in simple stuff?

MCPs

These MCPs are like sidekick gadgets, hooked up to my OpenCode to supercharge integrations.

Ah, Serena MCP — my secret sauce for making OpenCode a codebase ninja! It is like giving your AI X-ray vision for code.

It dives deep with precise symbol navigation (finding that sneaky variable’s birthplace), semantic search (hunting code like a pro detective), and AST-aware ops (that’s Abstract Syntax Tree for the uninitiated — basically, it understands code structure without getting tangled in the weeds).

Token-efficient, too, so it doesn’t gobble up your model budget while retrieving or tweaking code.

Jira MCP tames the wild beast of project management, diving into Jira issues and Confluence docs like a caffeinated intern. It’s local-flavored, launched with uvx mcp-atlassian, so you can query “What’s blocking sprint 42?” and get instant deets without logging into that soul-sucking web interface.

Say goodbye to ticket overload — OpenCode handles the drama while you sip coffee.

GitHub MCP is your ticket to repo raiding without leaving the terminal — think pulling issues, peeking at PRs, or cloning code snippets like a digital pirate.

No more tab-switching madness; just tell the AI “Fix that bug in my repo” and watch it yo-ho-ho the details. Arrr, efficiency!

Figma MCP brings the flair, fetching prototypes, frames, and feedback straight from your Figma designs. Perfect for when you’re mid-code and need to “Show me that UI mockup for the login page” — boom, OpenCode pulls it in for seamless dev-design mashups.

No more “Where’s that layer?” hunts; it’s like having a sassy stylist for your pixels, keeping everything fab and functional.

Skills

Enter skills.sh — the ultimate marketplace for reusable AI agent skills. These bad boys are like pre-packed procedural superpowers: modular chunks of know-how that teach your AI how to nail specific tasks, from debugging wizardry to React best practices, marketing copy, or even wild media generation.

After installing, swing by your skills/ folder (or wherever they land) and peek at the files — especially any SKILL.md or prompt configs. Some community skills are total Claude stans (looking at you, Anthropic-made ones) and hardcore “Claude” in the instructions or point to Claude-specific paths.

If OpenCode starts acting confused or throwing errors, manually edit those bits: swap “Claude” for generic terms to OpenCode-friendly phrasing.

Providers

Hong Kong’s got beef with Claude — total blackout due to the Supported Regions Policy. So, I pivoted to backup squads:

• Ollama, Opencode Zen for freebie models,
• Z.ai for GLM goodies, and
• xAI for Grok vibes.

No more “access denied” clownery — pure, barrier-free bliss.

Models

• Swapping GLM for Claude clone (Opus, Sonnet, you name it): Z.ai’s GLM-4.7 and GLM-4.6 step in like understudies stealing the show for reasoning and boss-level orchestration.
• Hooking up speedy, wallet-friendly xAI models for instant gratification: Stuff like grok-code-fast-1 nails codebase detective work and zippy jobs.
• Grabbing freebie gems from Zen and Ollama: kimi-k2.5-free and GLM-4.7-free, plus gemini-3-flash-preview, deliver no-cost heroes for writing, pondering, and general shenanigans, all wired per my repo magic.

Optimiztion

These hacks juggle model flavors and free rides to smash GLM’s pesky concurrency caps and squeeze every drop of value — like turning lemons into a lemonade empire.

• Mixing GLM variants to dodge rate-limit roadblocks: Spread the love across GLM-4.7 (cap 3), GLM-4.5 (cap 10), and pals to keep the party going without queues.
• Tossing in free models for flex appeal: Zen and Ollama’s zero-buck options mean you scale like a boss without the bill shock.

| Model Type     | Model Name    | Concurrency Limit |
|:---------------|:--------------|:------------------|
| Language Model | GLM-4.6       | 3                 |
| Language Model | GLM-4.7       | 3                 |
| Language Model | GLM-4.5       | 10                |
| Language Model | GLM-4.6V      | 10                |
| Language Model | GLM-4.7-Flash | 1                 |
| Language Model | GLM-4.5V      | 10                |
| Language Model | GLM-4.5-Air   | 5                 |
| Language Model | GLM-4.5-Flash | 2                 |

Here’s a cheeky table of GLM concurrency limits — because who doesn’t love a good cap chart?

Tips

How to Prompt Like a Pro and Hoard Those Tokens Like a Dragon? Just turn your prompt into JSON Gold!

Use a magical little tool like https://www.jsonprompt.it/ to transform your natural-language chaos into crisp, AI-friendly JSON prompts. You just paste your messy human instructions, and it spits out a structured JSON schema that tells the model exactly what to do, what format to output, and how to think step-by-step.

Why bother?

• LLMs love structure — it reduces hallucinations and gives you way more reliable results.
• It forces you to think clearly about what you actually want.
• Bonus: the output comes back in clean JSON, you can immediately parse or feed it into the next step.

It’s like giving the AI a treasure map instead of saying “uhh, the gold is somewhere over there-ish.”

These configurations, chilling in my GitHub, are my mad scientist tweaks for turbo performance, penny-pinching, and seamless development flow.