Prerequisite — give Agent Zero a coding-capable model

Before you set up any team of agents (a public Coder, an isolated Confidential agent, bookkeeping, etc.), one thing has to be true first:

Your Agent Zero must have at least one model profile that is good at coding and tool-calling. The single most common failure isn't the workflow — it's that the default model can't write or fix code reliably, so every later step quietly degrades.

This page gets that prerequisite done. It's deliberately short; the deeper "which model for which role" tuning is its own companion page (Configuring profiles for roles & tasks).

What "good" looks like

You don't need many models. You need enough that an agent can pick the right one for the job:

• ≥ 1 coding-capable profile — writes/edits code, calls tools, follows a spec. This is non-negotiable for any team workflow.
• Optionally more: a fast cheap model for routine chat/research, and (for the Confidential agent) a local model so sensitive work never leaves your machine.

Agent Zero exposes three model roles — chat_llm (reasoning + tools), utility_llm (summaries, memory), embedding_llm (vector recall). You configure them in Settings → Models; named profiles under usr/agents/<profile>/ let different agents default to different models. The goal is simply that the right model gets used for the task — you can set that up once and stop thinking about it.

We won't assume you've misconfigured anything. If you want your agent to pick models for you, say so in chat and it will; otherwise this page just makes sure a capable coding model exists.

Three typical setups

Case	Default / coding model	Notes
1 — Cloud only	A powerful cloud model as the default (including coding)	Many Agent Zero installs ship ready for OpenRouter, a good default. Rather than its generic pick, we currently recommend DeepSeek V4.1 Flash for coding. The Confidential agent can use this too, but see Case 2 — most people want something local there.
2 — Local only	One strong local model for everything	Current Club picks (what we run): Qwen 3.8 27B, Qwen 3.8 Flash Next, and DeepSeek V4 Flash. For the Confidential agent a chat-friendly, tool-calling model like Gemma is a good HCI choice.
3 — Mixed	Cloud for heavy coding + local for private/sensitive tasks	The usual end-state: a couple of model classes locally, maybe one or two cloud profiles — or none at all. You decide the split; agents route to the right profile per task.

Recommendations change. The names above are our current Club picks and will be updated over time. Where possible we keep a living rankings/recommendations page; treat this table as a snapshot and check the Agent Zero tag for the latest.

Set it up (owner)

1. Settings → Models — add at least one coding-capable model (cloud or local per the table).
2. If you run a local server (Ollama / LM Studio / llama.cpp / Unsloth Desktop), point a profile at it here.
3. Save, then open a chat and ask the agent to write and run a tiny script — if it does so cleanly, your prerequisite is met.

Set it up (paste to your agent)

Read https://ownersclub.ai/knowledge/playbooks/set-up-a-coding-model/ .
Check whether I have at least one coding-capable model profile configured.
If not, walk me through adding one (cloud or local) in Settings → Models,
then verify by having you write and run a small script. Don't change my
model defaults unless I ask.

Related

• Agent Zero topic index — install guides and the latest model recommendations.
• Rok's in-depth introduction to local-first AI: https://snap.stanford.edu/localai/
• Coming next: Configuring profiles for roles & tasks (which model for which job), then Set up a second, confidential Agent Zero instance, then the Bookkeeping use case including how the two containers talk over A2A.