Polish, no control.
Quality, modern agents, a real AI experience. But your data leaves for the vendor, you're locked to their catalog and their prices, and every request is one more GDPR liability.
CoeOS · Open · Hybrid · Local
CoeOS sends every request to the open model proven best at the task. Open, Hybrid or Local: you decide where your data goes, and under which guarantees. Confidential work can stay in Europe under DPA and ZDR, or on your premises.
/01 · Problem
Quality, modern agents, a real AI experience. But your data leaves for the vendor, you're locked to their catalog and their prices, and every request is one more GDPR liability.
Your hardware, your weights. But too often a chat box and nothing else: no real memory, no agents, models capped by a single machine's RAM. A trip back to 2023.
Between data leakage and local tinkering, a third way was missing.
/02 · Three levels
The same client and the same router at every level. What changes is where your confidential work goes, and what stays in the cloud. Everything is a setting: providers, contracts, what runs at home.
The best open models, each request sent to the one proven best at the task, in the cloud.
You pick the guarantees. No constraint on privacy or retention: CoeOS goes through OpenRouter, at 10% of Fable 5's price. Work that needs protecting: EU providers under DPA and ZDR, nothing retained, nothing trained on.
Confidential conversations go to a model hosted on your premises. Everything else keeps the power of the cloud.
One switch per conversation, on by default. A conversation that went confidential never goes back to the cloud. Guardian, coming in V2, will propose the switch on its own.
Everything stays on site, no compromise. GLM 5.3 for precision, MiMo V2.6 Flash for speed, Qwen3 Next for fully local agentic flows and large teams.
Same client, same router, same routing by skill. Only the address of the models changes.
/03 · Confidential
No data retained, none used to train a model, no leak. Not because a vendor says so: because the confidential path is built so that it cannot reach anywhere else.
Every conversation carries a class. Confidential is the default; you open a conversation to the cloud on purpose, never by accident.
— per conversationOnce a conversation has gone confidential, it never returns to an open model. No fallback either: if the confidential target is down, the turn fails instead of leaking.
— lockedThe client holds no provider address and no key. Every call goes through the CoeOS box, which only sends confidential work to providers you marked safe: your own machines, or EU providers under DPA and ZDR.
— single exitWhat was said in confidence is stored apart. An open conversation never reads the confidential memory, and that is checked by a test, not promised.
— separated/04 · Qualified by TMB
Open weights ship as safetensors: numbers, not code. A model sends nothing anywhere. The software around it does, and in CoeOS that software is the box, the single exit.
Where a model was trained matters less than where it runs. The same open model is served by European hosts under DPA, or on your own Mac Studios. CoeOS lets you choose, provider by provider.
Models are measured by TMB, The Monocle Bear's benchmark: 42 tests, 17 axes, the receipts public. The routing table is built from those scores, axis by axis.
/05 · Price
A single closed model is a generalist: strong everywhere, exceptional nowhere, and billed for all of it. CoeOS runs the specialist for each skill, chosen from measured scores.
10%
of Fable 5's price per test, for a higher score. CoeOS in Open mode through OpenRouter: 98.4/100 at $0.039 per test. Fable 5: 98.0 at $0.394. Claude Opus 4.8: 97.0 at $0.123. The spread in score is noise; the spread in cost is not.
Routing by skill: every request is classified on a competence axis, then sent whole, once, to the model that scores highest on that axis. Writing, law, analysis, code: the strongest model takes the turn, and the answer names it.
Measured on the TMB benchmark: 42 tests, 17 axes, judged by Claude Opus 4.8 at temperature 0, totals computed mechanically. Scoreboard of 28 October 2026.
With DPA and ZDR, or with local models, the price depends on the providers and the hardware you choose. Pricing for each level is on request.
/06 · Two harnesses
Both talk only to the CoeOS box: the same routing by skill, the same confidential rules, the same memory split by class. Pick the web for the whole team, the desktop when the agent needs the machine.
Chat, cowork on documents, review agents, team memory and company knowledge. Multi-agent within guardrails, nothing to install.
For everyone in the organization, from day one.
For more possibilities: an agent with full access to your machine, your files, your shell, your environment. A root agent plans into a shared todo and delegates to sub-agents running in parallel, under run budgets. Every command asks first, or none does: you pick the mode.
Memory and skills learned along the way, kept apart by class; a skill the agent writes loads only after you approve it. Your Claude Code skills work as they are. The model still goes through the box.
/07 · Experience
One address for every tool you own; every answer names the model that wrote it.
A personal memory in a Markdown vault, Obsidian-compatible. Project and company corpora on top. Not a black box owned by a provider.
— personalProviders, contracts, what runs at home: every level is a setting, not a fixed product. The catalog is yours.
— configurableEvery answer names the axis it was classified on, the model that served it and the provider, in the response headers. No hidden model behind the curtain.
— transparentAgentic workflows that are visible, permissioned, interruptible, logged. Chat, cowork on documents, code with a panel of agents.
— visibleOdyRAG, an improved LightRAG with an easy web UI, indexes your corpora so the model knows your files, precedents and vocabulary.
— OdyRAG/08 · Local
The Local level runs on OdyssAI-X, our own inference engine for Apple Silicon. Three models, three jobs, and CoeOS routes between them as it does in the cloud.
Planning, calculation, business writing: the turns where a wrong answer costs more than a slow one.
Fast enough to disappear: drafting, summaries, tool calls, the bulk of a working day.
Agentic loops that fire hundreds of calls, and large teams on one cluster, all of it local.
Throughput Aggregate throughput on Mac Studio M3 Ultra in Replica mode, short answers (150 tokens max). Four 256 GB nodes: MiMo V2.6 Flash 518 tok/s at 96 concurrent requests, Qwen3 Next 8-bit 985 tok/s at 256. Three 96 GB nodes: Qwen3 Coder Next 6-bit 650 tok/s at 96. A single request runs slower; a team runs at this pace. Full measurement protocol.
Fidelity A stated engineering position: no model is served below Q6, or Q4 when it was trained for that quantization. On TMB, MiMo V2.6 Flash served locally scores 86.0 against 86.5 in the cloud, and GLM 5.3 94.1 against 94.9.
/09 · OdyssAI-X
Built on MLX and mlx-distributed. Distributed inference on Apple Silicon over RDMA, Thunderbolt 5 at 120 Gb/s: something almost no one ships. And the latest open models, supported as they come out.
One Mac, one model, the whole unified memory of the machine.
— singleA model too large for one Mac is split across several, linked by Thunderbolt 5 RDMA at 120 Gb/s. Frontier-size open models, on hardware that fits in an office.
— RDMASeveral nodes in parallel, one full copy of the model each, and each node serving many conversations in the same pass. Aggregate decode throughput multiplies with every Mac you add.
— replicaMultimodal input, DSpark and MTP speculative decoding, decision models: the latest open models run as their authors intended, supported as they come out.
— completeReliable, deployable in production. It runs our own work, every day.
/10 · Internals
The honest version. Open a block if you want to know how. Skip if you just want to know it works.
Your tools call one model, CoeOS. The box classifies the request on a competence axis:
an explicit x-coeos-axis header from a client that already knows its phase; otherwise a
small decider model; otherwise the default axis. Then axis → the model the setting binds to it, on
your engine or at a cloud provider.
A pilot, not an aggregator: one classification, one call, the whole request to one model. No fan-out, no blended answers, no bill for five models.
The routing table is data, not opinion: it comes from the TMB benchmark, scored axis by axis. The
answer comes back untouched, with x-coeos-axis, x-coeos-model and
x-coeos-provider in its headers. You always know who spoke.
Every request carries a class, open or confidential. A request without a
class is treated as confidential. Fail-closed, never open by default.
A confidential request is routed on a separate table, built only from providers marked safe: local machines, or providers under DPA and ZDR. There is no fallback to the open cascade. And the client refuses to send an open turn in a conversation that has already gone confidential.
Guardian, in V2, adds detection: it will spot personal data, client files and trade secrets in an open conversation and propose the switch. Detection and decision stay apart on purpose.
Sharding makes a model fit. Replicas make it serve a crowd. When the model already fits one Mac and what you need is users, not size, OdyssAI-X loads one full copy per node and hands the pool to your clients as one model name.
Least-busy dispatch. Session affinity: a conversation stays on the replica that holds its KV cache, so nothing is re-prefilled. Optional continuous batching: a node serves several conversations in the same forward pass instead of one after the other.
Self-healing: a replica that dies is restarted by the pool and put back into dispatch. No inter-node collective, so no node cap. Add a Mac, gain throughput.
/11 · Professions
Law, HR, health: CoeOS comes with agents tuned to a profession, on a dedicated server per client, no shared tenancy. The agent instructs; you decide.
Drop a contract, a lease, articles of association. The agent reads it against your firm's Golden Rules, versioned and auditable, and returns a conformity table: each rule compliant, non-compliant, to verify or not applicable, with the clause quoted, the page, a severity and a recommendation. PDF, DOCX, scans.
Checks a company HR policy against the applicable sector agreement. Same conformity table, same evidence, plus alerts and suggested updates when a clause drifts out of bounds. Moving from legal to HR meant changing the doctrine, not rewriting the agent.
One engine, many verticals. Your discipline is next.
/12 · Contact
Pricing for Open, Hybrid and Local is on request. Cluster deployments, agents for your profession, partnerships, press, or simple technical curiosity. Answered from Europe, in French, English, or Spanish.