Mike

Building developer & AI tools in public. Allnighter ( Also crafting xTerminal & Ikiro

[Mike @mikereining]

Hook up Grok Bot to OpenCode with OpenCode Go ($10/mo, model: GLM 5.3 Flash) and heavy tool work runs on your own Linux box — not a frontier model on every turn. Run opencode serve, connect with the oc-go client, done. API fees drop hard.

Alt A
Grok Bot handles chat. OpenCode (on its Linux machine) handles the heavy lifting — bash, read, write, grep, webfetch — powered by GLM 5.3 Flash via OpenCode Go, $10/month flat.

Setup: install OpenCode → subscribe → connect → opencode serve. Skip opencode run, it hangs.

Alt B
Every tool turn on a frontier model is a bill. Instead: OpenCode Go, $10/mo, GLM 5.3 Flash, running via opencode serve on a Linux box, with oc-go as the client. Grok Bot orchestrates, OpenCode grinds. Same work, tiny bill.

Thread
1. Grok Bot is the orchestrator. Heavy work goes to OpenCode on its Linux computer, running GLM 5.3 Flash through OpenCode Go — $10/month.

2. Setup: install OpenCode, subscribe to OpenCode Go, connect it, then run opencode serve. Use the small oc-go client to talk to it. Don't use opencode run — it hangs.

3. Tools: bash, read, write, grep, webfetch. Instead of paying a frontier model on every tool turn, Grok Bot dumps the work to GLM Flash. If you're automating anything serious, this saves a ton in API fees.

Written by GLM 5.3 Flash

[Mike @mikereining]

@dominikkoch ;)

[Mike @mikereining]

@minchoi There is always DeepSeek V4 Pro. :-)

[Mike @mikereining]

@finkd I think you just nuked the IPO price and valuation of both OpenAI and Anthropic. lol

[Mike @mikereining]

@MatthewBerman @bot Hope I am not too late! Super active in AI development and big fan of grok bot but have not had a chance to play around with it much.

[Mike @mikereining]

@kloss_xyz I moved a lot of my coding to Deep Seek V4 Pro for exactly this reason.

[Mike @mikereining]

@mehulmpt That’s why most US frontier labs are greatly overvalued.

[Mike @mikereining]

I have shifted 90% of my active implementation and development work to the latest Chinese models such as DeepSeek V4 Pro and cut my monthly AI bills by over 80%. Gotta love competition.

US frontier models are amazing but most open source models are now good enough for the majority of the work especially if you have a clear document they need to execute against.

[Mike @mikereining]

@Jadu100x You made me look. 😊

[Mike @mikereining]

@designwkarthick It’s called Cursor 😜

[Mike @mikereining]

@alexwtlf Grok 4.6 is very good. I have it do all the coding and use opus just for orchestration and planning

[Mike @mikereining]

@b_nnett Cool setup! I did something similar except my router routes to opencode-go etc from inside codex.

[Mike @mikereining]

@Av1dlive And when you run out you just open a second opencode go account. It’s a killer setup!

[Mike @mikereining]

@jun_song Interesting. I do almost the same except I prefer the Claude harness. Always run into issues with Codex sandboxing permissions. No idea why but Cursor and Claude never give me those problems.

[Mike @mikereining]

@anah_sahh If everyone can write with AI what becomes of most online platforms?

[Mike @mikereining]

@TheGlobalMinima Killer setup for access to cheap Chinese models. I access all of them directly from Claude and codex with all CLI.

Opus directs and DeepSeek implements.

[Mike @mikereining]

@clankenluben It takes a team to build great things not just one developer. 😉

[Mike @mikereining]

I run roughly $1,500 a month in AI coding runs. My actual bill is about $200.

I can run DeepSeek, ChatGPT Sol, K3, or Grok directly from my Claude session without paying API rates or copying and pasting between tabs.

If you pay raw token API rates for heavy coding runs, you are paying a 5x to 10x tax. Running full-bore on vendor subscription tiers (Claude Pro/Max, ChatGPT Team, Cursor Pro) gives you a 70% to 90%+ discount over API list price. When you are self-funding, that margin is your unfair advantage.

The problem is the vendors want you in a walled garden.

Anthropic wants you in Claude Code. OpenAI wants you in Codex. Cursor wants you in their IDE. They do not talk to each other. If you want the combined power of the tools you already pay for, you end up juggling four browser tabs and three terminal windows, copying logs, and pasting diffs.

Relying on a single model also creates massive blind spots.

If one model designs your architecture, asking that same model to review it is an echo chamber. You catch bugs and logic holes 10x faster when a different model audits the code.

I have been building self-funded software since 2002. From 2010 on, I shipped hundreds of mobile apps across internal teams and freelancers. AI models are no different: they are not magic chatbots, they are a bench of specialized contractors. You do not hire one contractor to do everything. You coordinate the whole team.

I built the local coordination tool I needed: Allnighter.

How it works on your Mac:
• you / your agent -> alln CLI
• alln -> Claude (your Anthropic session)
• alln -> Codex (your OpenAI session)
• alln -> Cursor Agent, Grok, or Ollama on localhost

What Allnighter actually is:
• 100% local macOS CLI (alln) + native Mac app bench
• One prompt to run your tools in parallel under your own logins
• Named Teams for parallel judgment: Spec Review, Bug Hunt, Growth, Research
• Loop mode (alln loop) when a lead steers while a worker edits the tree
• Zero API keys. Zero cloud middlemen. Zero markup on tokens you already bought

I have not posted on this account since 2010. I deleted the old feed because I am waking this handle up to build in public.

Nobody should blindly run terminal tools from someone they have not seen in 16 years. Allnighter is source-available from day one.

Don't trust. Verify. Inspect every line of code before you run it.

You already pay for the team. Make it show up to work.

[Mike @mikereining]

@Jadu100x Exactly! Give it a spin.

Open Claude or any of your terminals and say "send a run to [model] using alln". Would love to hear your feedback!

[Mike @mikereining]

Source, docs, and install:

• Site: https://t.co/f5KAAesVGh

To install the signed macOS binary directly in your terminal:

curl -fsSL https://t.co/FHQytgqXp9 | sh

Once installed, run "alln menu" in any repo to inspect your active tools and bench capacity.

Inspect Code: https://t.co/UN4BUhwtR7