Faster development. Code stays in-house.
OpenCode with open-weights. Senior review systematic, not a bottleneck.
Development with Artificial Intelligence (AI) for Spryker and B2B platform teams. OpenCode as model-independent development agent, open-weights models (Llama, Mistral) on your own GPU hardware or at DACH hosts. Code map of the repo, review agents, governance hooks from day one. Senior review becomes a fixed process — not a bottleneck, and no code shipped to US cloud.
Three sovereignty anchors.
All on-prem-capable.
Cursor requires US cloud for premium features. GitHub Copilot runs on Microsoft cloud without an air-gap option. Anyone who needs code sovereignty — and that is significantly more in DACH industry than three years ago — needs a different stack.
OpenCode (sst/opencode) is a terminal-based engineering agent that communicates with any backend via an OpenAI-compatible API: Ollama locally, vLLM in a cluster, Anthropic, OpenAI, or Google Vertex as a cloud option. Self-hosting means fully egress-free operation. Cursor and GitHub Copilot have no comparable air-gap mode — that is the structural difference for companies with code sovereignty requirements.
Source: sst/opencode GitHub · OpenAI-compatible API specLlama 3.3 70B Instruct (Meta, December 2024) reaches performance close to Llama 3.1 405B at one sixth of the compute requirement. Int4-quantized it runs on a single H100 80GB at 30–80 tokens/s in single-user operation. Mixtral 8x22B (Apache 2.0) is fully commercial without special agreement — no license overhead for internal use. DACH GPU hosting providers: AD IT Systems, IONOS AI Model Hub, Hetzner GPU, Open Telekom Cloud, StackIT.
Source: Meta Llama 3.3 70B Instruct (12/2024) · Mixtral 8x22B Apache 2.0 (Mistral AI)Cloud frontier APIs currently cost USD 2-3 per million input tokens at Anthropic and OpenAI. From ~50–150 million tokens/month, on-prem hardware amortizes against cloud API costs — depending on the GPU leasing model and team size. Below that, cloud egress with a DLP layer can still be economically viable. Sovereignty remains possible then — via DLP egress filter, not hardware.
Source: Anthropic + OpenAI Pricing API (Q1 2025) · On-prem GPU calculation H100/A100 DACH
Model-agnostic.
Codebase-specific.
All four building blocks are model-agnostic: they work with Llama, Mistral, Anthropic, or OpenAI. What makes them codebase-specific are your conventions, your reference documents, your decisions.
Skills
Codified workflows as executable instruction sets: which steps in which order, which conventions apply, which anti-patterns to avoid. Skills are Markdown files — versionable, reviewable, model-agnostic. Anyone wanting to fight convention drift needs documented skills, not just a better model.
Workflows
The Workflow-Contract Research-Plan-Execute corresponds to the ReAct pattern from the ML literature — explicitly documented and traceable by every engineer. Phase separation prevents an agent from hallucinating instead of researching. Subagent patterns for parallel tasks are part of the architecture, not the vendor feature list.
Memory
Persistence across sessions: architecture decisions, known anti-patterns, codebase-specific conventions as reference documents. The agent remembers relevant facts without starting from scratch every time. Memory files are local — no external vector database as default, no egress for context data.
DLP Egress
Microsoft Presidio (MIT-licensed): regex + spaCy NER checks every prompt and output before cloud transmission. Customer IDs, internal endpoints, credentials — masked or blocked before they leave the network. Presidio runs locally as middleware, no external service, no egress for the DLP check itself.
Three patterns that slow engineering teams.
Three situations from active engagements — what keeps engineering teams from using AI-assisted workflows productively.
-
Pattern 01 · Code Egress Blocker
Legal blocks Cursor and Copilot. No pilot gets started.
Cursor requires US cloud for premium features; GitHub Copilot runs on Microsoft cloud without an air-gap. KRITIS operators, defense suppliers, companies with US customer data obligations cannot use these tools productively because code and context are transmitted to external servers. OpenCode + on-prem inference is the direct answer: same workflow quality, no egress by default.
- Consequence
- Engineering team works without AI support, competitive disadvantage grows
- Frame-Reset
- OpenCode self-host + open-weights: full functionality, no cloud dependency
-
Pattern 02 · Convention Drift
The agent does not know your architecture and invents its own.
Generic coding assistants generate code that is syntactically correct but convention-foreign — wrong namespaces, ignored abstraction layers, non-existent patterns. This creates review overhead that cancels the time savings. Skill architecture documents your conventions as executable references that the agent reads with every task. Drift becomes measurable, not just felt.
- Consequence
- Review overhead eats the productivity gain
- Frame-Reset
- Codebase-specific skills as convention anchor for every agent call
-
Pattern 03 · Senior Bottleneck
All complex tasks hang on one or two seniors.
Architecture decisions, complex refactorings, debugging in legacy areas — these tasks always land with the same people. Skill architecture codifies senior knowledge as executable workflows: the research phase clarifies the context, the plan phase sketches the solution, the execute phase implements per convention. Junior team members can handle more complex tasks; seniors review instead of implement.
- Consequence
- Knowledge silo, scaling ceiling, burnout risk for key people
- Frame-Reset
- Explicit workflows as portable senior knowledge in the team
An engineering agent is only as good as the conventions you give it. Without skill architecture it is expensive autocomplete. With it, it becomes a multiplier.
Five questions, five straight answers.
- "We are not allowed to send code to external providers. What is our option?"
- OpenCode + on-prem inference (Ollama or vLLM with Llama 3.3 70B or Mixtral 8x22B). Code and context never leave your network. Alternatively: DLP egress via Microsoft Presidio as middleware before cloud API calls — sensitive material is masked or blocked before transmission.
- "What is the difference between a skill and a prompt?"
- A prompt is one-time. A skill is a versioned, codebase-specific workflow description: phases, conventions, anti-patterns, reference documents. The agent reads the skill at every execution — like an engineer who knows the onboarding document. Portable across models, auditable, extensible.
- "How long until the pilot team is productive?"
- Phase 1 is scoped: codebase skill map, OpenCode setup with local model, first productive skills in the pilot team. Terms are mandate-specific. No time promises here — the depth and complexity of your codebase determine the effort.
- "Does on-prem hardware pay off for us?"
- Break-even is at ~50–150 million tokens/month against cloud frontier APIs (USD 2-3 per million input tokens). Below that, cloud egress with a DLP layer can be more economical. DACH GPU hosting providers (AD IT Systems, IONOS, Hetzner, StackIT) are often cheaper than US cloud at comparable GPU capacity. We calculate this concretely once we know your use case.
- "Can we switch to a different model later?"
- Yes. Skill architecture and workflows are model-agnostic. Switching models is a configuration task, not a reimplementation. Anyone using Llama 3.3 70B today and switching to a better open-weights model tomorrow keeps all skills unchanged.
Workbench, loops, handover. Three principles.
We build with you, not for you. Engagements are workbenches, not theatre. Delivery in short loops. Clean handover, then your team carries it alone.
Method in detailDeep platform expertise.
25 years in IT, 14 of them in B2B commerce. Architecture mandates with large enterprises, building and steering distributed expert teams, vendor-neutral project rescue. Focus areas: platform architecture, project rescue, team operations.
Chris Zepernick
Code sovereignty is not the exception — it is the starting point.
Anyone setting up engineering AI without an egress concept is building in a dependency that Legal will need to break later. OpenCode, open-weights, and DLP egress are not an academic exercise — they are the architecture choice that builds sovereignty in from day 1 without permanently blocking cloud options. The model is configurable. The architecture decision is not.
How we workCoding agent without cloud lock — does that fit your setup?
Code sovereignty blocker, convention drift, or senior bottleneck: a few minutes on the phone and we clarify which lever pulls first for you.
