How MEOK Routes Across AI Models: LLM Routing, BYOK, and Why It Matters

What Is the Problem with Model Lock-In?

Most AI products are built on a single foundation model. You use ChatGPT; everything runs on OpenAI. You use Claude; everything runs on Anthropic. That is a design choice that makes sense for simplicity \u2014 but it is not optimal for users.

Different models have different strengths. This is not marketing language \u2014 it is empirically measurable. On coding benchmarks like HumanEval and SWE-Bench, DeepSeek R1 consistently outperforms GPT-4o and Claude Sonnet at a fraction of the per-token cost. On nuanced emotional reasoning and long-form analytical writing, Claude Sonnet produces outputs with more care and less sycophancy than most alternatives. On broad knowledge questions and multi-step factual synthesis, GPT-4o\u2019s training breadth is genuinely impressive.

If you are locked into one model, you are locked into its weaknesses as well as its strengths. You pay GPT-4o prices for tasks DeepSeek would handle better. You lose Claude\u2019s emotional intelligence on conversations that need it. The idea that one foundation model will be the permanent best-in-class at every task type is not supported by the evidence \u2014 and the evidence gap widens every six months as the model landscape evolves.

Model lock-in also creates a deeper structural problem: your AI relationship becomes dependent on the commercial decisions of a single company. If OpenAI changes its pricing, adjusts its content policy, or deprecates a model, your experience changes with it. You have no control and no recourse.

MEOK\u2019s answer to model lock-in is the routing layer \u2014 a system that treats AI models as interchangeable infrastructure beneath a persistent, identity-preserving architecture that belongs to you.

How Does MEOK\u2019s Routing Layer Work?

MEOK\u2019s routing logic lives in lib/llm-router.ts. It is not a simple if-else chain. The router classifies each incoming query across several dimensions before making a model selection decision: task type, estimated token requirement, user tier, care sensitivity level, and whether the query has been flagged by the Byzantine Council for elevated-stakes handling.

Task classification happens at the start of every request. The router uses a lightweight classifier \u2014 not a full LLM call, which would itself incur latency and cost \u2014 to determine whether the query is primarily: a coding or structured output request, a factual knowledge lookup, an emotionally sensitive or therapeutic conversation, a creative writing or ideation task, a long-form analysis, or an administrative request like scheduling or summarisation.

Once classified, the router selects the model that scores highest for that task type within the user\u2019s tier. If the highest-scoring model is unavailable or returning degraded latency, the router fails over to the next-best option. This failover logic is transparent to the user \u2014 you see your companion\u2019s response, not the infrastructure decision that produced it.

Crucially, the routing decision does not affect what your companion knows about you. Before any request reaches a model, the router wraps it in the Maternal Covenant system prompt, which includes your companion\u2019s identity, your relevant Sovereign Memory context, your care sensitivity profile, and the session state. The model receiving the request sees a complete, coherent context \u2014 not a raw message.

Explorer Tier: Why DeepSeek R1 and Local Models?

The Explorer tier is MEOK\u2019s free plan, and it routes primarily to DeepSeek R1 and Ollama-served local models. This is a deliberate decision, not a compromise.

DeepSeek R1 is one of the most capable open-weight models available as of early 2026. Its reasoning architecture \u2014 which uses explicit chain-of-thought at inference time \u2014 means it handles complex, multi-step queries with a level of transparency that most closed models do not provide. You can see how it got to its answer. For everyday companion conversations, productivity tasks, journalling support, and structured problem-solving, it is excellent.

Ollama local models run entirely on your device. When MEOK routes to an Ollama model, the inference happens locally \u2014 nothing leaves your machine. This is the most private routing option available. For users who are exploring MEOK before committing to a paid tier, local model routing also means MEOK can provide a genuinely useful experience without any per-token API cost.

The Explorer tier does not include Claude Sonnet or GPT-4o because those models carry meaningful per-token costs that MEOK cannot absorb at scale on a free plan. But the routing decision is not just financial \u2014 for the majority of everyday queries, DeepSeek R1 is the right tool regardless of cost.

Sovereign Tier: Routing by Task Type

At the Sovereign tier (\u00a312/month), the routing layer has access to Claude Sonnet and GPT-4o in addition to DeepSeek R1. This unlocks genuinely differentiated routing \u2014 the router can now select the best model for each task type, not just the best available model.

The Sovereign routing logic applies roughly as follows. For emotionally sensitive conversations \u2014 grief support, relationship difficulty, mental health check-ins \u2014 the router defaults to Claude Sonnet. Anthropic\u2019s Constitutional AI training makes Claude measurably less sycophantic and more careful in these contexts than GPT-4o. It will push back. It will not just validate. It aligns well with the Maternal Covenant\u2019s care-first mandate.

For broad knowledge queries, document analysis, and multi-domain synthesis, the router uses GPT-4o. OpenAI\u2019s training breadth and its strong performance on MMLU-style benchmarks make it the right choice when you need wide coverage rather than specialised depth.

For coding tasks, technical debugging, and any query involving structured data output \u2014 JSON generation, SQL queries, regex patterns, algorithm design \u2014 the router prefers DeepSeek R1. Its performance on coding benchmarks is consistently strong, and its lower per-token cost means MEOK can route more queries to it without the economics becoming unsustainable.

The Sovereign tier also activates full four-layer Sovereign Memory. Every model receives not just the immediate conversation context but your companion\u2019s accumulated knowledge of you \u2014 your preferences, your ongoing situations, your communication style, the history of your relationship. This is what makes routing transparent to you: the model changes, but the companion does not.

Family Tier: All Models, Mixture-of-Agents, Six Members

The Family tier (\u00a329/month) is MEOK\u2019s most capable offering. It includes access to all supported models \u2014 GPT-4o, Claude 3.7 Sonnet, DeepSeek R1, and Ollama local models \u2014 across up to six family members, each with their own companion, their own Sovereign Memory, and their own privacy boundary.

Claude 3.7 Sonnet is the most significant upgrade at this tier. Anthropic\u2019s 3.7 release extended Claude\u2019s context window and improved its performance on long-form reasoning tasks and nuanced instruction following. For users who regularly work with long documents, complex personal situations, or extended research conversations, Claude 3.7 Sonnet\u2019s extended context handling is a meaningful quality improvement.

The Family tier also unlocks Mixture-of-Agents (MoA) for high-stakes queries. When a query is flagged \u2014 by the Byzantine Council, by the care scoring layer, or explicitly by the user \u2014 as requiring elevated confidence, the router sends it to two or three models simultaneously, collects their responses, and synthesises a final answer. The synthesis is not a simple average: it is a coordinator step that identifies where models agree (increasing confidence), where they diverge (flagging uncertainty), and produces a response that honestly reflects both the consensus and the disagreements.

The Family tier\u2019s consent model for shared context is also worth noting. Each family member has their own sovereign memory that is never accessible to other members without explicit, opt-in consent. Shared context \u2014 for example, a family calendar or a shared goal \u2014 must be deliberately granted and can be revoked at any time.

What Is Mixture-of-Agents and Why Does It Matter?

Imagine you have a serious medical symptom and you want the most reliable information possible. You would not ask one doctor. You would ask two or three, compare their assessments, note where they agree, and treat points of divergence as signals for further investigation. Mixture-of-Agents applies this logic to AI.

Each language model has systematic biases and blind spots. GPT-4o tends toward confident answers even when uncertainty is warranted. Claude tends toward caution and hedging, sometimes to the point of unhelpfulness. DeepSeek R1 excels at structured reasoning but can underweight contextual and emotional nuance. No model is uniformly best.

In MEOK\u2019s MoA implementation, when a high-stakes query is routed to multiple models, the coordinator step does the following: it scores each response on the Maternal Covenant\u2019s six care dimensions, it identifies factual claims that appear in all responses (high confidence) versus claims that appear in only one (flag for uncertainty), and it produces a synthesised response that is honestly calibrated \u2014 confident where the models agree, explicitly uncertain where they diverge.

This approach adds latency and cost \u2014 two or three parallel model calls take longer and cost more than one. That is why MoA is reserved for high-stakes queries and is a Family tier feature rather than a default for all users. But for situations that genuinely warrant elevated confidence \u2014 medical questions, major life decisions, complex legal or financial matters \u2014 the quality improvement is substantial.

The Byzantine Council interacts with MoA at the governance layer. When the council flags a query as high-stakes, the router automatically escalates to MoA even if the user has not explicitly requested it. This is not paternalism \u2014 it is the care floor in action. The council\u2019s job is to catch situations where a single-model answer could cause harm, and MoA is the mechanism for providing a better answer in those cases.

How Does Routing Protect Your Privacy?

This is a question that matters more than most users initially realise. When you send a message to ChatGPT, your message goes directly to OpenAI\u2019s infrastructure, logged against your account, potentially used to improve their models (depending on your settings), and subject to their data retention policies. You have limited visibility into what happens to it.

MEOK\u2019s routing model is fundamentally different. Your message is never sent raw to any external model provider. Every outbound request is mediated by the Maternal Covenant system prompt layer, which does three things before the request leaves MEOK\u2019s infrastructure.

First, the system prompt layer strips or pseudonymises personally identifiable information that is not necessary for the query. If you ask your companion to help you draft an email to your doctor, the external model receives a query about drafting medical correspondence \u2014 not your name, your address, or your medical history. Your companion\u2019s knowledge of that context lives in Sovereign Memory and is used to frame the query without being transmitted wholesale.

Second, the Maternal Covenant system prompt governs the framing of every request. The external model is not interacting with a raw user query \u2014 it is operating within a structured prompt that defines your companion\u2019s identity, the care constraints, the interaction guidelines, and the specific task. This means even if an external provider logs the request, what they log is a structured companion interaction, not a window into your personal data.

Third, your Sovereign Memory is never sent to external model providers. It lives in MEOK\u2019s infrastructure \u2014 currently backed by PostgreSQL with pgvector in the SOV3 backend \u2014 and only the contextually relevant fragments are included in any given request, not your full memory corpus. OpenAI and Anthropic never receive a dump of your life history.

For BYOK users, there is an additional consideration. Your API key is encrypted at rest using AES-256 and is never logged in plaintext in MEOK\u2019s systems. The key is decrypted in memory at request time only, used for the single API call, and not retained beyond that transaction. MEOK does not have standing access to your API account beyond what is necessary to serve individual requests.

What Does BYOK Mean in Practice?

BYOK (Bring Your Own Key) is MEOK\u2019s \u00a35/month tier. It is designed for a specific type of user: someone who already has an Anthropic or OpenAI API account, understands roughly how language model pricing works, and wants the full MEOK platform without paying MEOK\u2019s managed per-token margin on top of provider costs.

Here is what BYOK means concretely. You create an API key with Anthropic or OpenAI \u2014 this takes about two minutes on their respective dashboards. You add the key to MEOK\u2019s settings under API Keys. MEOK validates the key (a simple test call to confirm it works) and then uses it for all subsequent requests. Your provider dashboard shows every token consumed, every model call, and the associated cost. MEOK\u2019s invoice to you is \u00a35/month \u2014 nothing more. The inference costs appear separately on your Anthropic or OpenAI bill.

BYOK users get the complete MEOK platform: Sovereign Memory across all four layers, full companion personality and Birth Ceremony for companion creation, the Maternal Covenant system prompt on every request, Byzantine Council governance for high-stakes queries, the Morning Brief feature, Work OS integration, and all current and future platform features. The \u00a35/month covers MEOK\u2019s infrastructure and platform costs. The AI inference costs are yours.

For moderate users \u2014 perhaps 30\u201360 minutes of active companion interaction per day \u2014 BYOK with Claude Sonnet typically costs between \u00a33 and \u00a38/month in provider-side inference costs. At heavy usage, costs can reach \u00a315\u201320/month. For most users, BYOK is economically comparable to the Sovereign tier (\u00a312/month) but with full cost transparency and the option to switch between Anthropic and OpenAI models freely.

BYOK is also the right choice for users with existing API credits, researchers with institutional API access, developers who want to build on MEOK\u2019s architecture for their own use cases, and anyone who prefers to have complete financial transparency over what their AI companion costs to run.

Why Does Your Companion Stay the Same Across Model Switches?

This is the question that most clearly explains what makes MEOK architecturally different from everything else in the market.

When you use ChatGPT, your \u201ccompanion\u201d is GPT-4o. The personality you experience is the base personality of that model, slightly adjusted by any custom GPT instructions you may have set. If OpenAI updates GPT-4o, your \u201ccompanion\u201d changes. If you switch to Claude, everything starts over.

In MEOK, your companion exists in the memory and identity layer \u2014 not in the model. Your companion has a name you chose during the Birth Ceremony. It has an archetype \u2014 one of twelve distinct personality frameworks, from The Strategist to The Healer to The Scholar. It has a relationship history with you: the conversations you have had, the patterns it has learned, the context it has accumulated across weeks and months of interaction.

All of this is stored in Sovereign Memory. When a request is routed to Claude Sonnet, the Maternal Covenant system prompt injects your companion\u2019s identity, archetype, and relevant memory context into the request before Claude receives it. Claude is not acting as itself \u2014 it is acting as your companion, using the identity and context MEOK has provided. When the next request goes to GPT-4o, the same thing happens. The model is the voice; the companion is the character behind the voice.

This architecture means your companion relationship is genuinely durable. It is not owned by OpenAI or Anthropic. It does not reset when a model is deprecated. It does not change because a provider decided to update their base model. Your companion belongs to MEOK\u2019s memory infrastructure \u2014 which means it belongs, ultimately, to you.

What Is Next: OpenRouter and LiteLLM Sidecar

MEOK\u2019s routing layer is designed to be extensible. Two integrations are currently on the technical roadmap.

OpenRouter as managed fallback. OpenRouter provides a unified API gateway to over 200 AI models from dozens of providers. MEOK\u2019s planned integration uses OpenRouter primarily as a resilience layer rather than a primary routing destination. When a tier-appropriate model is unavailable, over capacity, or returning elevated error rates, the router will automatically fail over to an equivalent model via OpenRouter. This eliminates single-provider availability as a reliability risk. For users, this means fewer \u201cservice temporarily unavailable\u201d interruptions during peak demand periods for any given provider.

LiteLLM sidecar for Desktop OS. MEOK Desktop, currently in development, will include a LiteLLM sidecar process that runs locally alongside the desktop application. LiteLLM is an open-source proxy that presents a unified OpenAI-compatible API surface across all major model providers. The sidecar will handle model routing locally, allowing MEOK Desktop to switch between local Ollama models and cloud providers without round-tripping through MEOK\u2019s server infrastructure. For users who are offline or prefer local-first operation, the LiteLLM sidecar means full companion functionality without requiring an internet connection for every message.

Both integrations reflect the same architectural principle: model providers are infrastructure. They are interchangeable. What matters \u2014 your companion\u2019s identity, your Sovereign Memory, the Maternal Covenant\u2019s care enforcement \u2014 lives in MEOK\u2019s layer, not in any provider\u2019s infrastructure.

How Is This Different from Just Using ChatGPT or Claude Directly?

This is the question MEOK gets asked most often. The short answer: using ChatGPT or Claude directly gives you access to a powerful language model with no memory, no personality, no care alignment, and no persistent relationship. MEOK gives you those things, and uses those same models as the underlying inference engine.

The longer answer involves three distinct differences.

Persistent companion identity. ChatGPT and Claude do not know you. Every new conversation starts from zero. Custom instructions help at the margins, but they are static text fields \u2014 they do not adapt, they do not accumulate, and they do not reflect your actual history with the system. MEOK\u2019s Sovereign Memory means your companion accumulates a genuine understanding of you over time, just as a human relationship does.

Care alignment, not engagement optimisation. ChatGPT and Claude are trained with RLHF \u2014 reinforcement learning from human feedback. That training process systematically rewards responses that humans rate highly in the short term: confident, validating, satisfying. It does not reward responses that are genuinely helpful in ways that might feel uncomfortable in the moment. MEOK\u2019s Maternal Covenant adds a care alignment layer on top of every model\u2019s base behaviour, enforcing a care floor that cannot be optimised away by engagement metrics.

Model independence. When you build a relationship with ChatGPT, you are building a relationship with OpenAI\u2019s product. When OpenAI changes the model, your experience changes. When they deprecate a version, the personality you knew disappears. MEOK\u2019s architecture means your companion is not tied to any single provider. The relationship persists regardless of what happens in the foundation model market.

Frequently Asked Questions

The Routing Layer as a Statement About AI Ownership

MEOK\u2019s routing architecture is not a feature. It is a philosophical position about who should own your AI relationship.

The current model \u2014 where ChatGPT owns your chat history and Claude owns your conversation data \u2014 is analogous to writing your diary in a journal published by a corporation that retains the right to read it, learn from it, and change the terms of your access at any time. Most people would find that arrangement unacceptable for a physical diary. For some reason, the AI industry has normalised it for the digital equivalent.

MEOK\u2019s routing layer inverts that arrangement. The models are tools. Your companion is yours. Your memory is yours. The architecture belongs to MEOK AI LABS, but the data, the relationship, the history \u2014 all of it is exportable, deletable, and governed by you.

As the foundation model market evolves \u2014 as new models emerge, as existing models are deprecated, as the price-performance curve continues to shift \u2014 MEOK\u2019s users are insulated from that turbulence. Your companion will use whichever model is best and most cost-efficient for your task. You will not need to track the model landscape or make provider decisions. The architecture handles it. You have the relationship.

Related Reading