Inspiration
Enterprise supplier onboarding on SAP Ariba is broken. Every new supplier requires 100+ manual touchpoints, takes 30–120 days, and bottlenecks at 5–10 concurrent onboardings. The root cause isn't complexity — it's that every step is coordination overhead: emails, status checks, EDI validation, connectivity tests, guide generation. The work is structured. The bottleneck is human bandwidth. WebMCP finally gives agents a browser-native way to take that load off.
What it does
AribaFlow is a WebMCP-powered procurement integration portal where AI agents and procurement teams work side by side to onboard suppliers in 7–14 days instead of 120.
Agents handle the 80% that's repetitive:
- EDI document validation — schema checks, ANID mismatch detection, field mapping verification
- CIG connectivity testing — AS2/VAN/HTTPS endpoint verification, certificate status
- Integration guide generation — auto-draft from test results, ready for human review
- Onboarding checklist orchestration — tracks 13-step SI/SIFL workflow, surfaces blockers
- Supplier readiness scoring — ranked queue so integration teams prioritise the right suppliers
Humans handle the 20% that matters: governance, sign-off, and exceptions.
All agent interactions happen through WebMCP tools registered on the portal — no backend API keys exposed, no separate agent infrastructure, just a website that agents can actually use.
How we built it
The stack: Next.js 14 (App Router) on Vercel, WebMCP Imperative API for tool registration, React frontend with a live onboarding dashboard, and a mock Ariba/CIG backend that mirrors real document flows (PO 850/OrderRequest, ASN 856, Invoice 810).
WebMCP tools are registered at page load via navigator.tools.register(). Each tool maps 1:1 to
a procurement workflow action — structured inputs, structured JSON output, designed so an LLM
can chain them without ambiguity.
The domain model comes from real Ariba integration experience: CIG adapter patterns, AS2 connectivity, cXML/X12 document types, and the SI/SIFL/NEL role separation that governs every enterprise onboarding project.
Challenges we ran into
- WebMCP origin isolation — any
document.domainusage breaks the tools API. Had to audit every third-party embed and iframe to enforce origin isolation across the portal. - Tool schema design — JSON Schema for EDI validation inputs is non-trivial. Too loose and agents hallucinate field values. Too strict and valid documents fail. Iterated with the WebMCP Inspector extension until tool calls were reliable.
- Demo vs. reality — real Ariba systems require corporate VPN and supplier ANIDs. Built a faithful mock layer that exercises the same error paths (ANID_MISMATCH, MAP_FIELD_MISSING, CERT_EXPIRED) so the demo reflects genuine failure modes, not happy-path theatre.
What we learned
WebMCP's browser-native model is genuinely different from server-side MCP. The agent and the human share the same page state — which means the portal UI and the agent's tool responses have to be in sync. That constraint forced better product design: every tool response updates visible UI state, so the human always knows exactly what the agent did and why.
What's next
- Live Ariba Network API integration (with OAuth, not mock)
- Declarative WebMCP annotations on the supplier self-service forms
- Multi-buyer tenancy so integration service providers can run parallel onboarding waves
- Compliance audit trail: every agent action logged with timestamp, tool input, and output
Built With
- as2
- b2b-integration
- chrome-devtools-protocol
- cig
- cxml
- edi-x12
- gemini
- json-schema
- next.js
- node.js
- openai
- procurement-automation
- react
- rest-api
- sap-ariba
- supply
- tailwind-css
- typescript
- vercel
- webmcp
- webmcp-inspector
Log in or sign up for Devpost to join the conversation.