"The Codexes Family — Human-Centric AI for the People Who Own the Problem"
"Four local-first workbenches for domain experts: Domain, Prompt, Local, and Photo. Shared philosophy. Different jobs."
The Codexes Family
Human-Centric AI starts with one belief: the domain expert's knowledge is the model. Everything else — the algorithm, the maths, the code — is plumbing.
The Codexes are four desktop workbenches built on a single conviction: the people who understand the problem should be the people who build the solution. Each Codex focuses on a different shape of expertise — predicting, instructing, talking, seeing — but all four share the same architecture, the same philosophy, and the same uncompromising answer to one question:
Where does your data go?
Nowhere. It stays on your laptop. There is no server to send it to, because there is no server.
The shared manifesto
| Local-first. | Every Codex runs entirely on your hardware. No cloud, no account, no telemetry. Your data never leaves the device. |
| Citizen-driven. | Domain experts lead. The Codex translates plain language into structured, auditable artefacts — you don't translate yourself into Python. |
| White-box by default. | Every model, every prompt, every rule, every classification is readable. Defensible to a regulator, an auditor, or a student. |
| Open formats. | AGPL-3.0 source. Open file formats. Your work is portable, exportable, and yours forever. |
| AI Act-ready. | Architecture aligns with EU AI Act Article 13 transparency, GDPR Article 5 data minimisation, and NIST AC-7. Compliance is the design, not a checklist. |
The four Codexes
🧭 Domain Codex — for predictions you must defend
The workbench for domain experts in regulated industries.
You know the underwriting rule, the clinical signal, the fraud pattern, the sensor reading, the policy outcome. Domain Codex turns that intuition into a working machine-learning model: describe what you want in plain English to the bundled Scribe assistant, and the platform builds, validates, and explains a model on your laptop. Supervised, unsupervised, causal discovery, rules, Apex AutoML, Formulas, Recipe Replay — the full stack, locally.
Who it's for: underwriters · clinicians · factory engineers · compliance officers · analysts · policy advisors · AI architects.
Editions: Browser (try it now, no install) · Lite · Community · Professional · Enterprise.
✍️ Prompt Codex — for prompts that survive production
The workbench for treating prompts as engineering artefacts.
If your organisation puts language models into customer-facing decisions, the prompt is the spec. Prompt Codex is where prompts become first-class: versioned, evaluated against a golden set, A/B tested across model providers, and shipped through an auditable pipeline. Roll a prompt back the way you'd roll back a deployment. Diff two prompts the way you'd diff two pull requests. Run a regression suite the way you'd run unit tests.
Who it's for: prompt engineers · platform teams · LLM operators · compliance officers in companies that deploy generative AI to customers.
What it does locally: prompt library + versioning · golden-set evaluation · regression tests · provider-agnostic adapter (OpenAI / Anthropic / local Ollama) · audit trail of every prompt change and every test run.
🏠 Local Codex — for AI that never leaves the building
The local LLM platform for organisations that cannot use the cloud.
Hospitals, banks, ministries, defence contractors, law firms — anyone whose data legally, ethically, or commercially cannot be sent to a third-party provider. Local Codex installs, manages, and serves local language models (Llama, Qwen, Mistral, your own fine-tunes) on infrastructure you control. With prompt routing, basic evaluation, and the same audit trail as the other Codexes.
Who it's for: IT leads at regulated organisations · CISOs · sovereignty- conscious public sector · research institutions with sensitive cohorts.
What it does locally: model registry + version pinning · Ollama-based inference · prompt routing between models · embedding + retrieval (RAG over your private corpus) · access logs anchored to a hash-chained audit trail.
📷 Photo Codex — for images you cannot upload
The local computer-vision workbench for domain experts who work with pictures.
Medical images that cannot leave the hospital network. Industrial inspection photos that cannot leave the factory floor. Archive scans that cannot leave the museum. Insurance damage photos that cannot leave the underwriter's laptop. Photo Codex lets you organise labelled image datasets, train and validate a classifier or detector, and run inference on new images — all locally. The same explainability principles as Domain Codex (saliency maps, decision rationale, confidence intervals) applied to vision.
Who it's for: radiologists · pathologists · quality engineers · archivists · loss adjusters · field biologists · anyone whose images carry legal weight.
What it does locally: dataset organisation + labelling · classifier and detector training (PyTorch, ONNX export) · per-image explanation overlays · batch inference with confidence thresholds · model card generation for audit.
What they share — the Codex architecture
Each Codex is a desktop application that follows the same blueprint.
The same core powers all four. Only the artefacts differ:
| Codex | The artefact |
|---|---|
| Domain | A trained model with its training corpus, lineage, and explanation |
| Prompt | A versioned prompt with its golden set and regression history |
| Local | A served LLM with its routing rules, embedding index, and access log |
| Photo | A trained vision model with its labelled dataset and saliency overlays |
What flows between them is shared: hash-chained audit logs, recipe replay, sovereignty by architecture, AGPL-3.0 source.