Pivert's Blog

Reading Time: 7 minutes

Executive Summary

Professional-grade AI coding for 15–20 hours per week now lands closer to €20/month than €30/month, stays online when one provider hiccups, and does not need a €150/month inference budget. The June 2026 update to our OpenCode + Oh My OpenAgent stack is built on three simple decisions:

• Use the free Zen flash model first. Bulk agents now start on opencode/deepseek-v4-flash-free, fall back to the paid Zen flash model only if needed, then hop to ollama-cloud/deepseek-v4-flash.
• Stop burning GPT on Oracle. The biggest OpenCode surprise was not the free models; it was GPT-5.4 usage. Oracle now runs on ollama-cloud/deepseek-v4-pro, so GPT spend is limited to Momus and Hephaestus fallback.
• Keep Level 3 as the default backbone. ollama-cloud/kimi-k2.7-code and ollama-cloud/glm-5.2 still carry most of the stack, with V4 Pro reserved for reasoning-heavy work.

The result: a stack that is cheaper, still resilient, and much less likely to quietly leak money through OpenCode pay-per-use calls.

What Is Oh My OpenAgent?

OpenCode is an open-source AI coding framework that supports dozens of providers, MCP servers, and plugins. By default it is a single-model chat: you pick a model and talk to it.

Oh My OpenAgent is the plugin that turns OpenCode into a multi-agent development system. It adds specialised agents, each with independent model routing and ordered fallback chains. You do not have to pick one model for everything; the plugin routes each task to the agent and model that fit it best.

The Agent Catalog

Agent	What it does	When to use it	Model used here
Sisyphus	Orchestrator. Receives every message, plans, and delegates to the right specialist.	Default for almost everything.	ollama-cloud/kimi-k2.7-code
Prometheus	Strategic planner. Interviews you, then writes a verified plan before any code is touched.	When the idea is vague or the change is critical.	ollama-cloud/deepseek-v4-pro
Atlas	Todo orchestrator. Executes approved plans across sessions.	After Prometheus produces a plan.	ollama-cloud/glm-5.2
Metis	Pre-planning analyst. Catches hidden constraints and ambiguities.	Automatically invoked by Prometheus.	ollama-cloud/kimi-k2.7-code
Momus	Ruthless plan reviewer. Validates clarity, verifiability, and completeness.	The one place we still accept GPT-5.4 cost.	opencode/gpt-5.4
Oracle	Read-only architecture consultant. High-IQ reasoning for unfamiliar patterns and tradeoffs.	@oracle for security, architecture, and hard debugging.	ollama-cloud/deepseek-v4-pro
Librarian	Documentation and OSS search.	When you need to know how library X does Y.	opencode/deepseek-v4-flash-free
Explore	Fast codebase grep. Pattern discovery and “where is X?” searches.	When you need files, symbols, or conventions fast.	opencode/deepseek-v4-flash-free
Multimodal-Looker	Vision analyst. Reads screenshots, PDFs, diagrams.	Any visual input.	opencode/gemini-3.1-pro
Sisyphus-Junior	Focused implementation worker. Writes one assigned unit and cannot re-delegate.	Spawned automatically during execution.	ollama-cloud/kimi-k2.7-code
Hephaestus	GPT-native deep worker. Give it a goal, not a recipe.	Rare deep cross-domain reasoning.	opencode/gpt-5.5

Read-only agents (Oracle, Librarian, Explore, Multimodal-Looker) cannot write or edit. Momus cannot write or edit. Sisyphus-Junior cannot re-delegate. These restrictions keep each agent in its lane.

How To Invoke Agents Directly

Oh My OpenAgent registers every agent into OpenCode’s @-mention picker. You can override the default routing explicitly.

@-mentions

Shortcut	Routes to	Typical use
@oracle	Oracle	Architecture review, security, hard debugging
@librarian	Librarian	Docs, examples, library internals
@explore	Explore	Find code patterns, references, conventions
@plan "…"	Prometheus	Create a structured plan before coding

All registered agents also appear in the picker (@sisyphus, @hephaestus, @prometheus, @atlas, @momus, @metis, @sisyphus-junior, @multimodal-looker).

Keyword Modes

Keyword	Effect
ultrawork / ulw	Full parallel-agents execution mode
search / find	Web and documentation search focus
analyze / investigate / audit / deep-dive	Deep context-gathering analysis
hyperplan / hpp	Adversarial plan review
hpp ulw / ulw hpp	Hyperplan + ultrawork combined

Slash Commands

Command	Effect
/start-work	Activates Atlas on the latest Prometheus plan
/plan	Switches to Prometheus for structured planning
/refactor	LSP + AST-grep assisted refactoring
/review-work	Spawns parallel QA agents
/handoff	Generates a context summary for a new session
/ulw-loop	Self-referential loop in ultrawork mode

The Cost Reality: GPT Was the Problem, Not Zen Itself

The expensive surprise in real usage was not deepseek-v4-flash on Zen. It was GPT-5.4.

In practice, Oracle was one of the main sources of OpenCode pay-per-use spend. A single heavy day could burn around $12, and a single month could quietly land around $20 just on GPT-5.4. That is too much for a tool that is supposed to feel economical.

So the new rule is blunt:

• Oracle moves to ollama-cloud/deepseek-v4-pro
• Deep work moves to ollama-cloud/deepseek-v4-pro
• GPT-5.4 stays only on Momus
• Hephaestus keeps GPT because the plugin is GPT-native there

This is the real cost win. The free flash model saves pennies and parallelises nicely. Removing GPT-5.4 from everyday reasoning agents saves dollars.

—

New Design Principle: Cheap First for Bulk, Ollama First for Reasoning

The earlier version of this stack overused the phrase “cross-provider redundancy first”. That is still true in spirit, but the actual winning policy is simpler:

• Bulk agents use the free Zen flash model first
• If Zen free rate-limits, use paid Zen flash second
• If Zen is unavailable, hop to Ollama Cloud flash
• Reasoning-heavy agents live on Ollama Cloud by default

Examples:

Primary	First Fallback	Second Fallback	Why
opencode/deepseek-v4-flash-free	opencode/deepseek-v4-flash	ollama-cloud/deepseek-v4-flash	Cheapest possible bulk path
ollama-cloud/kimi-k2.7-code	opencode/kimi-k2.6	ollama-cloud/deepseek-v4-pro	Cheap orchestration with a Zen safety net
ollama-cloud/deepseek-v4-pro	ollama-cloud/kimi-k2.7-code	opencode/gemini-3.1-pro	GPT-free reasoning path
ollama-cloud/glm-5.2	ollama-cloud/deepseek-v4-pro	ollama-cloud/kimi-k2.7-code	Atlas stays fully on Ollama

The important point is not philosophical purity. It is that the stack should stay online without silently shifting half the bill onto OpenCode GPT calls.

—

Agents vs Categories: Two Different Routing Layers

Oh My OpenAgent exposes two independent routing layers. Les confondre rend la config difficile à lire, donc les voici séparées.

Agents

Les agents sont des personnages adressables que tu peux mentionner avec @ ou que Sisyphus appelle explicitement. Chacun a sa propre configuration de modèle. Ils sont le “qui”.

Agent	What it does	When to use it	Model used here
Sisyphus	Orchestrator. Receives every message, plans, delegates.	Default for almost everything.	ollama-cloud/kimi-k2.7-code
Prometheus	Strategic planner. Writes a verified plan before code.	Quand l’idée est floue ou critique.	ollama-cloud/deepseek-v4-pro
Atlas	Todo orchestrator. Executes approved plans across sessions.	Après un plan Prometheus.	ollama-cloud/glm-5.2
Metis	Pre-planning analyst. Catches ambiguities and constraints.	Appelé automatiquement par Prometheus.	ollama-cloud/kimi-k2.7-code
Momus	Ruthless plan reviewer. Validates plans.	The one place we still accept GPT-5.4 cost.	opencode/gpt-5.4
Oracle	Read-only architecture consultant.	@oracle for security, architecture, hard debugging.	ollama-cloud/deepseek-v4-pro
Librarian	Documentation and OSS search.	Recherche de docs et d’exemples.	opencode/deepseek-v4-flash-free
Explore	Fast codebase grep and pattern search.	Trouver du code dans le repo.	opencode/deepseek-v4-flash-free
Multimodal-Looker	Vision analyst. Reads screenshots, PDFs, diagrams.	Tâche visuelle.	opencode/gemini-3.1-pro
Sisyphus-Junior	Focused implementation worker. Cannot re-delegate.	Spawné automatiquement en exécution.	ollama-cloud/kimi-k2.7-code
Hephaestus	GPT-native deep worker. Give it a goal, not a recipe.	Raisonnement profond cross-domain rare.	opencode/gpt-5.5

Read-only agents (Oracle, Librarian, Explore, Multimodal-Looker) cannot write or edit. Momus cannot write or edit. Sisyphus-Junior cannot re-delegate.

Categories

Les catégories ne sont pas des agents. Ce sont des profils de modèle utilisés quand un agent délègue via task(category="..."). Sisyphus ou Prometheus peuvent dire *“ce morceau de travail est une tâche visuelle”* et laisser le plugin choisir le modèle adapté sans hardcoder d’agent.

Category	Primary Model	Fallback Strategy	When to use
quick	opencode/deepseek-v4-flash-free	paid Zen flash, then Ollama flash	Trivial tasks
unspecified-low	opencode/deepseek-v4-flash-free	paid Zen flash, then Ollama flash	Moderate tasks
unspecified-high	ollama-cloud/kimi-k2.7-code high	opencode/kimi-k2.6, then V4 Pro	Complex multi-file tasks
deep	ollama-cloud/deepseek-v4-pro	Kimi, then Gemini 3.1 Pro high	Premium technical work
visual-engineering	opencode/gemini-3.5-flash	Kimi, V4 Pro, Gemini 3.1 Pro high	UI/UX, styling
ultrabrain	ollama-cloud/deepseek-v4-pro max	Kimi, then Gemini 3.1 Pro high	Maximum reasoning
writing	opencode/deepseek-v4-flash-free	paid Zen flash, then Ollama flash	Documentation
infrastructure	ollama-cloud/deepseek-v4-pro	Kimi	IaC, Terraform, K8s, Ansible

Routing automatique interne

Sisyphus ne choisit pas le modèle final à la main. Il choisit un canal :

• @oracle ou @librarian → agent nommé avec son propre modèle.
• task(category="visual-engineering", ...) → category, le plugin route vers le modèle de cette catégorie.
• task(subagent_type="explore", ...) → agent nommé.

Il n’y a pas de correspondance forcée entre un agent et une catégorie. Par exemple visual-engineering est rarement appelée par l’agent multimodal-looker ; c’est Sisyphus qui l’utilise quand il identifie un besoin UI/UX. De même, librarian est un agent de recherche, mais une simple tâche d’écriture de doc ira plutôt en category: writing.

L’intérêt de séparer les deux couches :

• Agents = contrôle explicite et adressable. Tu sais exactement qui lit ton message.
• Categories = contrôle implicite par type de travail. Sisyphus n’a pas besoin de créer un agent pour chaque micro-tâche.
• Coût par type de travail, pas par personnage. Le bulk écriture coûte le même prix flash qu’une exploration, sans forcer un routage vers Librarian.

Updated Agent Assignments (June 2026)

Agent	Primary Model	Fallback Strategy	Rationale
Sisyphus	ollama-cloud/kimi-k2.7-code	opencode/kimi-k2.6, then V4 Pro	Every message. Keep it Level 3 and cheap.
Oracle	ollama-cloud/deepseek-v4-pro max	Kimi, then Gemini 3.1 Pro high	The main GPT-cost saver.
Momus	opencode/gpt-5.4 high	V4 Pro, Kimi, Gemini 3.1 Pro high	Worth paying for a ruthless critic.
Prometheus	ollama-cloud/deepseek-v4-pro	Kimi, then Gemini 3.1 Pro high	Planning likes V4 Pro; no GPT required.
Plan	ollama-cloud/deepseek-v4-pro	Kimi, then Gemini 3.1 Pro high	Same logic as Prometheus.
Metis	ollama-cloud/kimi-k2.7-code	opencode/kimi-k2.6, then V4 Pro	Cheap pre-planning analysis.
Sisyphus-Junior	ollama-cloud/kimi-k2.7-code	opencode/kimi-k2.6, then V4 Pro	Bulk implementation stays Level 3.
Librarian	opencode/deepseek-v4-flash-free	paid Zen flash, then Ollama flash	Free-first bulk search.
Explore	opencode/deepseek-v4-flash-free	paid Zen flash, then Ollama flash	Free-first codebase grep.
Atlas	ollama-cloud/glm-5.2	V4 Pro, then Kimi	Large-context coordination at Level 3 price.
Multimodal-Looker	opencode/gemini-3.1-pro	Kimi	Vision stays on Gemini.
Hephaestus	opencode/gpt-5.5 medium	GPT-5.4 medium, V4 Pro, Kimi	GPT-native plugin path; leave it rare.

Updated Category Assignments (June 2026)

Category	Primary Model	Fallback Strategy	Rationale
quick	opencode/deepseek-v4-flash-free	paid Zen flash, then Ollama flash	Free-first trivial tasks.
unspecified-low	opencode/deepseek-v4-flash-free	paid Zen flash, then Ollama flash	Free-first moderate tasks.
unspecified-high	ollama-cloud/kimi-k2.7-code high	opencode/kimi-k2.6, then V4 Pro	Complex multi-file tasks.
deep	ollama-cloud/deepseek-v4-pro	Kimi, then Gemini 3.1 Pro high	Premium technical work without GPT.
visual-engineering	opencode/gemini-3.5-flash	Kimi, V4 Pro, Gemini 3.1 Pro high	UI/UX stays on Gemini.
ultrabrain	ollama-cloud/deepseek-v4-pro max	Kimi, then Gemini 3.1 Pro high	Maximum reasoning.
writing	opencode/deepseek-v4-flash-free	paid Zen flash, then Ollama flash	Free-first documentation.
infrastructure	ollama-cloud/deepseek-v4-pro	Kimi	IaC still benefits from V4 Pro.

What disappeared from the active mix:

• GPT-5.4 as a routine fallback all over the stack
• ollama-cloud/glm-5.1 and opencode/glm-5.1
• unnecessary paid Zen flash calls for bulk agents

—

Why This Is Better Than the Previous “Optimised” Version

The previous revision was already smarter than the default plugin setup, but it still had one leak: too many routes eventually ended at opencode/gpt-5.4.

That was the expensive mistake.

Three concrete improvements now:

• Oracle no longer burns GPT spend. The biggest hidden OpenCode cost is gone.
• Bulk agents finally use the actual free model first. If Zen gives you deepseek-v4-flash-free, use it.
• GPT is now explicit, not accidental. You pay for Momus because it is worth it. You keep Hephaestus GPT-native because the plugin expects it. Everything else is pushed back to Ollama Cloud or free Zen.

—

Updated Monthly Cost Breakdown

Component	Cost	Notes
Ollama Cloud Pro (annual plan)	€15/month	Flat-rate, generous GPU quotas
OpenCode Zen pay-per-use	~€3–7/month	Mostly Momus, occasional Hephaestus, Gemini vision/UI calls
TOTAL	~€18–22/month	15–20h/week heavy coding

This is the first version of the stack that feels honestly cheap. The saving does not come from turning everything into a weak free model. It comes from being disciplined about where GPT is allowed to exist.

—

Privacy and Data Residency: Still the Same Strong Story

The new models do not change the privacy picture. Kimi K2.7-code, GLM-5.2, and DeepSeek V4 are open-weight models pulled from HuggingFace and hosted on Ollama Cloud’s US/EU NVIDIA infrastructure. Data does not flow to Chinese API endpoints. Ollama’s terms remain zero-logging, zero-retention.

When the primary model is opencode/..., inference runs through OpenCode Zen’s infrastructure, which is a separate provider. The redundancy design therefore also means sensitive data is not tied to a single provider’s pipeline.

—

The Full Configuration

The complete oh-my-openagent.jsonc keeps every model ID verified against:

• https://opencode.ai/zen/v1/models for OpenCode Zen models
• https://ollama.com/search?c=cloud for Ollama Cloud models

Conclusion

AI-assisted development does not need a €150/month inference budget, and it definitely does not need accidental GPT bills.

The June 2026 update is the first version of this stack that feels both resilient and honestly cheap:

• Kimi K2.7-code for orchestration and implementation
• GLM-5.2 for large-context coordination
• DeepSeek V4 Flash Free for bulk search and documentation
• DeepSeek V4 Pro for heavy reasoning without OpenCode GPT spend
• GPT-5.4 / GPT-5.5 reserved for the rare places where they are actually justified

That is the real trick: not “never use premium models”, but “stop using them by accident”.

—

Resources

• Ollama Cloud Pro pricing: https://ollama.com/pricing
• OpenCode Zen pricing: https://opencode.ai/docs/zen
• OpenCode Zen model list: https://opencode.ai/zen/v1/models
• Ollama Cloud models: https://ollama.com/search?c=cloud

—

Author Note: This configuration is refined through daily use. The June 2026 revision was driven by three observations from production: (1) Level 4 quota burn is real, (2) the free Zen flash model is good enough for bulk work, and (3) GPT-5.4 was the main OpenCode cost leak, so it had to be pushed out of the default reasoning path.