Technical Walkthrough
This page is for technical readers who want the operating model without jumping straight into implementation details.
Runtime Map
The platform is split into distinct runtime surfaces:
| Runtime | Port | Responsibility |
|---|---|---|
| Gateway API | 8002 | Provider-compatible proxy, governance chain, health, runner admin routes |
| B2B API | 8001 | Dashboard/admin API for authenticated product operations |
| Dashboard | 3001 | Operator console for keys, logs, costs, approvals, health, and runners |
| Docs | 3003 | Developer documentation site |
The Core Flow
Client SDK / app
-> Gateway
-> Auth + scopes
-> Provider resolution
-> Governance chain
-> Upstream provider call
-> Usage recording
-> Dashboard, billing, and monitoring surfacesGateway Request Path
The gateway is designed to accept provider-style requests while adding organization-aware governance.
Main stages
- Authenticate Curate-Me gateway key.
- Resolve org and key context.
- Enforce scopes such as
gateway:write. - Parse request body and requested model.
- Resolve provider and concrete upstream model.
- Resolve provider credentials from headers, stored secrets, or environment fallback.
- Run governance checks.
- Proxy the request.
- Record cost and usage.
Governance checks
The exact implementation evolves, but the mental model stays stable:
- plan enforcement
- rate limiting
- cost and budget checks
- PII or secret scanning
- model allowlists
- HITL approval gating
Managed Runners
The runner system exists for work that needs an execution environment rather than a single model response.
Examples:
- browser automation
- file manipulation
- shell commands
- scheduled workflows
- multi-step agent jobs with persistent state
Runner operating model
Dashboard / API
-> Runner control plane
-> Template / policy / credentials resolution
-> Provider or host allocation
-> Session execution
-> Files, logs, activity, cost, and lifecycle eventsImportant runner concepts
| Concept | Purpose |
|---|---|
| Templates | Reusable runner configurations for repeated jobs |
| Tool profiles | Limit what a runner can access |
| Fleet | Grouped runner inventory and activity |
| Your Machines | Connect your own hardware (laptop, server, or VM) to the control plane |
| Cloud deploy | One-click managed provisioning on platform infrastructure |
| Intent router | Convert natural-language job intent into template-backed execution |
Dashboard Data Model
The dashboard consumes two broad categories of backend surfaces:
| Source | Typical usage |
|---|---|
| B2B API | auth, org settings, billing, dashboard widgets, admin actions |
| Gateway API | public health, provider health, gateway-specific operations, runner admin routes |
In practice that means the dashboard is not just a thin shell over one API. It bridges both the admin plane and the gateway plane.
Health, Logs, And Cost Flow
Gateway request
-> response completes
-> usage record created
-> Redis accumulators updated
-> MongoDB usage/audit persisted
-> dashboard log viewer and billing sync consume the resulting dataWhat this enables:
- request-level logs
- cost rollups
- budget enforcement
- provider health views
- operator investigation after incidents
Audience-Specific Reading Paths
Use this if you are explaining the product internally or to customers.
Plain-English VersionUse this for rollout planning across engineering, finance, support, and leadership.
By RoleUse this when you want the longer-form technical architecture page.
Deep Architecture