Broke up with Claude Code: open-weight models at about two cents a task

Move the default agent to OpenCode on open-weight models via a LiteLLM gateway. Homelab developer who left Claude Code as his default.

Share on 𝕏 ↗ Compare with… Read the source ↗ Full teardown →

what to steal

Put every model call behind one gateway with per-app budgets, and force a test run before 'done'.

Our takeaway from the source, not a quote.

Drawn from this card's sourced fields. Amber = older model named by the source. Rose = cut or replaced.

joekarlsson-opencode-litellm.toml

# Move the default agent to OpenCode on open-weight models via a LiteLLM gateway
# source: https://www.joekarlsson.com/blog/breaking-up-with-claude-code/

[rig]
who     = "Homelab developer who left Claude Code as his default"
author  = "Joe Karlsson"
date    = "2026-10"

harness = ["OpenCode", "Claude Code"]
models  = ["DeepSeek V4.1 Flash", "Claude (escalation for the hardest problems)"]

[keep]
  - Every model call goes through one LiteLLM gateway on a Proxmox cluster
  - Apps ask for aliases like cloud-coder / local-fast; the gateway picks the model
  - Per-app keys and budgets: OpenCode gets $20 a month
  - OpenCode plugin that sends the agent back once if it edits code and finishes without running anything
  - Fixed OpenCode so DeepSeek gets its earlier reasoning back
[cut]
  - Claude Code as default
  - Trusting leaderboards
  - Coding agent on homelab GPUs

thing	status	why (per source)
Claude Code as default	replaced	Claude kept as the escalation path
Trusting leaderboards	replaced	Tests models on his own two-task test; V4.1 Flash passed every run at about two cents a task
Coding agent on homelab GPUs	dropped	"my homelab GPUs can't run a coding agent"

Source: https://www.joekarlsson.com/blog/breaking-up-with-claude-code/ · checked 2026-10-14

Ticket in, PR out: four skills and a reviewer in CI

steal → Split the pipeline into ticket, plan, build and address-review skills.

Claude CodeOpenCode Claude Opus 5.5GPT-6 Astra GSD, Aider, tmux, Term

by rogs

~/rigs/vanja-opencode-deepseek-glm.toml2026-09-16

OpenCode on DeepSeek V4.1 Flash or GLM-5.3 instead of a native seat

steal → Point your harness at a cheaper model endpoint and test it on your own tasks.

OpenCodeClaude CodeCodex DeepSeek V4.1 FlashDeepSeek V4-Pro-0813 deepseek-coder model n

by Vanja Petreski

~/rigs/nadimtuhin-mac-controls-pc-r….toml2026-09-01

Laptop as control room, home PC as the engine

steal → Run agents on an always-on box behind one model gateway; reach it with zero open ports.

Claude CodeCodexPi Any model behind 9Rout

by Nadim Tuhin