☰ On this page
This layer does not shrink your agent's bill. Claude Max, Cursor, or whatever you code with keeps its own price. What moves is the work: reviewing, triaging and summarising happen on a cheap model inside Oprex, so what reaches the expensive agent is already clear. If your expensive agent is already doing both jobs, the honest answer is to switch this off — and that takes one click.
Where the two names came from
The split is older than Oprex. You hire an architect to decide what is to be built and to write it down precisely; you hire a handyman to build exactly that. The architect is expensive and you use little of it. The handyman is cheap and you use a lot of it. The whole saving depends on the architect's drawing being good enough that the handyman never has to guess.
Applied to software with AI: a strong model writes and repairs the Specification and the Requirement, and a cheap model does the repetitive work those documents describe. Get it the wrong way round — a cheap model deciding what to build, an expensive one fixing the mess — and you pay twice.
What they are, concretely
| Layer | What it does | Which model |
|---|---|---|
| Architect | Reads each Requirement and Specification for clarity, completeness and testability. On plans with gating review it can hold the save until the artefact is fixed; on the free plans it writes its findings onto the artefact and never blocks. | Your workspace's AI — the architect model if you set one, otherwise the plan's smart model. Deliberately allowed to be a small model: its job is to check, not to compose. |
| Handyman (Autopilot) | Takes a critical bug, reads the suspected code in your repository, drafts a fix, and opens a Pull Request for a human to review. | Your workspace's fast model. Disabled by default and gated by plan — see Autopilot. |
What actually spends your AI tokens
Every AI call Oprex makes is labelled, and the same labels appear in Billing → AI usage in your panel, so this table can be checked against real numbers rather than trusted.
| Label | When | What it is |
|---|---|---|
arsitek |
Automatic | Every Requirement and Specification you save is read before (or just after) it is stored. This is the only layer that spends tokens without anyone clicking. |
ai-fill |
On click | The "AI Fill" button on a description or acceptance-criteria field. |
balasan-komentar |
On click | A suggested reply on a comment thread. |
triase-tiket |
On click | Triage a helpdesk ticket into severity, type, and routing. |
ringkasan-tiket |
On click | Summarise a long ticket thread. |
cari-kb |
On click | Search the knowledge base in plain language. |
mcp-ai-chat |
On call | The oprex_ai_chat tool, called by your agent. |
deteksi-project |
On call | Detect which project an artifact belongs to. Off unless asked for. |
memory-facts |
On call | Turn a session into durable facts. |
autopilot |
Staff-run | The Handyman layer. Disabled by default; see below. |
uji-failover |
On click | The "Test chain" button in Settings › Integration. A few tokens, on purpose. |
One of these is automatic. Everything else waits for a click or a tool call. So the entire question of "is Oprex quietly spending my credit?" comes down to a single setting.
Two setups. Pick the one you are actually in.
You already drive Oprex from a strong agent
You use Claude Code, Cursor, or a similar agent on a paid plan, connected over MCP. That agent already writes the Specification, already drafts the Requirement, and already repairs the code — it is both architect and handyman. A second, weaker model reviewing its work afterwards adds cost and a second opinion you did not ask for.
Set AI Architect review to "Off" in Settings → Integration. Nothing in Oprex will spend your provider's tokens unless you press a button. The traceability, the gates, the release flow — all of that is ordinary software and keeps working without any model.
You are budget-limited
Students, campuses, communities, small teams: no per-seat agent subscription, and the AI budget is whatever fits. This is what the layer was built for. Point Oprex at a cheap provider — DeepSeek, Groq, or an Ollama model on your own machine — and leave the Architect on.
The saving is real and it is specific: a cheap model catching "this requirement cannot be tested" costs a fraction of a cent, while the same discovery made later, by an expensive agent, costs a full context of reading code that should never have been written.
Where to change it, and where to check it
- Switch: Settings → Integration → AI Architect review — Off, Advisory (reviewed after saving, never blocks), or Gating (can hold the save). Your plan sets the ceiling; you choose at or below it.
- Status: the bar at the bottom of the panel says which layers are live right now, without opening settings.
- Spend: Billing → AI usage — tokens per feature, per model, per day, per person, with the wasted ones counted separately.
- Model: Settings → Integration → Model (architect) — leave it blank to follow the plan's smart model, or name a smaller one.
Free to start, and free to switch off
Bring your own key, or your own local model, or no model at all. Oprex is a lifecycle tool first; the AI layer is an option you control.
Get started free Open AI settings Ask us about your setup