# ADR 005: opencode

- HTML version: https://robbiepalmer.me/projects/homelab/adrs/005-opencode
- Project: Home Lab (https://robbiepalmer.me/projects/homelab.md)
- Status: Accepted
- Date: 2026-08-02

# Context

The hub's agent lineup is Claude Code, Codex, and Grok Build — three
proprietary agents, each tied to a commercial model family and subscription.
I want at least one open-source agent in the mix for the same reasons the
rest of the lab is built on open tooling: no account dependency, no per-token
surprise, and the freedom to point it at any model.

opencode is an open-source, terminal-based agent that is intentionally
model-agnostic — it can run against API providers or any OpenAI-compatible
endpoint. That's what matters here, because open weight is how you get to the
interesting models:

* **Free model access.** Providers such as DeepSeek (and Big Pickle) offer
  genuinely capable models at no cost, which a subscription-only lineup can't
  touch.
* **Open-weight frontier models as they land.** New frontier open-weight
  releases (e.g. Kimi K3) become usable through third-party providers without
  waiting for a commercial agent vendor to wire them in.

To be clear: I'm **not** planning to self-host models. Open weight is a good
direction, but local inference isn't — self-hostable models are too small and
weak, and the frontier open-weight models (Kimi K3 reportedly needs on the
order of $2M of GPUs to serve a single instance) are only runnable by
third-party providers anyway. Open weight means provider competition, not a
box in my house.

# Decision

Install **opencode** on the Mac mini alongside the proprietary agents.

opencode provides the model-agnostic fallback: if a provider's allowance is
exhausted or its pricing changes, opencode can be pointed at a different
model without changing the tool. It's also the way to use free and
open-weight models — DeepSeek, Big Pickle, Kimi K3, whatever comes next —
through whichever provider serves them.

# Alternatives

## Open source agents other than opencode (Aider, Continue)

* **Pros**: Aider and Continue are mature, well-known open-source options.
* **Cons**: opencode's model-agnostic design and clean `AGENTS.md`-style
  conventions fit the hub best.
* **Decision**: Rejected. opencode is the orchestrator-friendly choice.

## No open-source agent; rely on the three commercial ones

* **Pros**: Fewer tools to maintain.
* **Cons**: Leaves no model-agnostic fallback; all agent capacity depends on
  commercial accounts.
* **Decision**: Rejected. The open-source slot is deliberate — it's the
  resilience fallback for the whole agent stack.

# Consequences

### Pros

* **Model-agnostic fallback**: works against any OpenAI-compatible endpoint.
* **Free and open-weight models**: DeepSeek, Big Pickle, and frontier
  open-weight releases like Kimi K3 are a few lines of config away, through
  third-party providers.
* **Open source**: no account dependency, inspectable, and free.
* **Complements the commercial agents** as the fourth in the lineup, all
  driven from the same orchestration surface.

### Cons

* **Younger ecosystem**: fewer battle-tested integrations than Claude Code.
* **Another tool to configure and keep updated.**
* **Not the primary agent**: by design it's the fallback, so it gets less
  real-world exercise than the commercial agents.

---

Markdown index of this site: https://robbiepalmer.me/llms.txt