GR Work Tech · team onboarding

How the Brain works

Three things often get confused — the app, the engine, and the model. Once they are clear, most of our open questions answer themselves. No technical background needed.

draft for internal review · June 2026

The three layers

Most confusion comes from treating these as one thing. They are not.

Three stacked layers: model, runtime, app
Layer 1

The model

The raw intelligence — Claude, DeepSeek, a local model. It reasons and writes. It is interchangeable: you can swap it without changing anything else.

On its own it is a mind in a jar: it can think, but it cannot do.

Layer 2

The runtime

Claude Code — the engine that gives the model hands and memory. It reads and writes your files, runs tools, keeps the Brain's knowledge up to date.

Runtime and model work together. This is what makes the Brain an assistant, not just a chatbot.

Layer 3

The app

The web interface you actually use. Upload, voice, buttons, history, sharing. It is the room you meet the assistant in.

Each Brain runs in its own sealed box (Docker), so one client's data never touches another's.

A useful way to hold it: the model thinks, the runtime acts, the app lets you in. When something feels wrong, the first question is always which layer.

Why the Brain beats "just using Claude"

Comparing the Brain to a chat window is comparing an employee to a phone call.

When you open Claude or ChatGPT in a browser, you get a model in a chat box: smart, but with no memory of your business, no access to your files, no ability to act. Every conversation starts from zero.

The Brain is the model plus the runtime plus your accumulated knowledge. It remembers your candidates, your projects, your way of writing. It can open a document, draft an email in your voice, update its own records. That value lives in the runtime and the memory — and it stays even if we change the underlying model.

So "I want Claude" and "I want the Brain" are not in competition: the Brain can run on Claude, on DeepSeek, or on a local model. The model is a setting inside the product — so the answer to "bring Claude back" is simply to enable Claude as one of the dials.

Why we need the app, not the command line

The engine is powerful from a terminal. The business cannot live there.

Claude Code is, at heart, a command-line tool — excellent for a developer, hostile for everyone else. The web app exists because the things GR Work Tech actually needs are impossible or painful from a terminal:

The app is not a nicety. It is the surface on which the product becomes usable by a team and saleable to clients.

The model question

There is no single "best" model. There are trade-offs, and a plan.

The request"DeepSeek isn't as good as Claude for our writing — bring Claude back."
The realityA fair preference: Claude is strong for redactional work. The catch is that Claude is not currently enabled on this Brain — it needs credentials — so today it cannot actually be selected, even when it appears in the menu. That is a switch to flip, not a verdict on the product.
ModelQualityCostNotes
ClaudeTop for writinghighFavourable pricing is temporary; login/auth is painful, especially inside a web app.
DeepSeekReasoning ~ on par~nothingThe pragmatic default for most work.
Local (on our GPU)Lower ceilingfixedThe compliance fallback — runs on hardware we control.

The plan: OpenRouter as the dial

The fastest way to give everyone every model is OpenRouter: one integration, one key, access to Claude, GPT, DeepSeek and the rest, selectable per task. That is why it wasn't wired in earlier on its own — it is not free. OpenRouter adds roughly a 5% surcharge on top of each provider's price. Small in itself, but pure overhead — and it does not change the real cost driver, which is which model runs underneath.

What it costs

Real usage over 30 days, all three users aggregated, measured from both the chat database and the session logs:

Dialogue — chat database (what was actually said)975 messages · ~192K tokens
Processing — session logs (full context + cache)~971M tokens · ~17M fresh
Cost basis (what the engine actually bills)the ~971M, mostly cheap cache reads

One user drives almost all of it (heavy daily Claude use, large context re-read each turn); the other two are light. Benchmark: 3 Pro licences = $300/month. Each row below is the provider's most expensive model — the ceiling.

Provider — top modelReal usage / monthvs $300 (3 licences)
DeepSeek V4 Pro client default~$188under
DeepSeek V4 Flash light / internal~$1619× less
Grok 4.3 (xAI)~$260under
Gemini 2.5 Pro~$390over
Claude Sonnet 4.6~$6402× more
GPT-5.5~$9803× more
Claude Opus 4.8~$1,0653.5× more
Cost / month at real usage — each provider's top model DeepSeek and Grok stay under the 3-licence line; Claude & GPT shoot past it at heavy-user volume. DeepSeek Pro (client default) ≈ $188; Flash (light use) ≈ $16. $0$250 $500$750$1000 DeepSeek V4 Pro Grok 4.3 Gemini 2.5 Pro GPT-5.5 Claude Opus 4.8 $188 $260 $390 $980 $1,065 3 Pro licences · $300
reference (DeepSeek, Claude) other cloud (GPT, Gemini, Grok) +5% via OpenRouter — — $300 = 3 Pro licences (threshold). GPT/Gemini approx (cache pricing varies).

OpenRouter adds ~5% to any row (the teal sliver). Figures cost the measured token mix at each provider's rate; the dominant term is cached context — priced ~100× higher on Claude than on DeepSeek — which is what splits the chart.

The headline: one heavy user drives the whole cost. DeepSeek keeps grbrain under the licence line — ~$188 on Pro (the client default), ~$16 on Flash for light use. Metered Claude, at that volume, runs past $300 (~$640 Sonnet, ~$1,065 Opus). So DeepSeek is the production engine; Claude is a metered route used where its writing earns its keep; and a single user that heavy is the one case where a personal subscription would beat metered Claude — for that one person, on their own login.

The OAuth trap — why not just share a subscription?

A natural reflex: skip the API bill entirely and put a Claude subscription (Max, ~$100–200/mo) behind the platform. It looks free at the margin — and it is a closed door. A subscription authorizes one person, interactively, at their own terminal. Wiring its login token into the platform to serve many users is, in effect, reselling one personal subscription as a service — against Anthropic's terms, capped by a single shared rate limit, and a single point of failure the day they tighten it (which they are doing). It does not scale by construction. For a single heavy user it would be cheaper than metered Claude — but it stays blocked: you cannot share one login across the team, and DeepSeek is cheaper than both anyway.

The clean line is by usage mode, not by login mechanism:

There is no clean way to get "cheap unlimited Claude" for a multi-user service. Either you pay per token via the API, or it isn't Claude. The OAuth shortcut tried to dodge that economic fact — and that is exactly the door being shut.
So the roles settle out: OpenRouter is the flexibility layer — one key to try every model. DeepSeek is the default engine — Pro for real client work (~$188/mo), Flash for light/internal (~$16), both under the licence line. Claude is a metered route used where its writing earns the cost (at heavy-user volume it runs $640–1,065/mo, so use it deliberately, not as the default). The model is a dial — cost does constrain it here, so leave it on DeepSeek and turn to Claude for the jobs that justify it.

What a "skill" is

A capability you add — independent of which model is running.

LinkedIn scraping, Outlook inbox reading, candidate presentation, job descriptions: these are skills. They live in the runtime layer and work the same regardless of the model behind them. Adding LinkedIn is not a model question — it is installing a skill. Outlook is the same kind of work, to be built.

This matters because it untangles a frequent knot: "the AI can't do X" is usually not the model being weak — it is a skill not yet installed.

Compliance: why a local model must exist anyway

Recruitment runs on personal data. Sovereignty is not optional, long-term.

A vault built from data blocks — data kept under control

Even if a cloud model is excellent and cheap, GR Work Tech handles candidate data — names, histories, contacts — under GDPR. For the most sensitive work, the data should be able to stay on hardware we control. That is why we keep researching a fully local fallback, accepting a quality ceiling in exchange for the data never leaving.

It only becomes economically sensible above a certain client base — the full analysis is here:

Going deeper — the two analyses

This page is the overview. Two companion documents go one level down — one on the architecture, one on the infrastructure.

The frame in one line: the company brain is distributed in the architecture (it emerges) and can be sovereign in the infrastructure (it runs on hardware you control).

What this means for us, now

Short version, so nobody waits on a misunderstanding.

The product is the Brain — memory plus runtime plus app. The model is a setting inside it. Once the team holds that distinction, the roadmap reads clearly.