What Is Subport, Really? And Why Doesn't Claude Code + MCPs Make It Redundant?
A sober look at where the real value lives in an AI-native support platform.
The One-Sentence TLDR
Subport is an operating system for AI-driven customer support — not a chatbot, not a wrapper around an LLM, but a multi-tenant runtime that orchestrates persistent conversations, human handoffs, SLA-bound tickets, knowledge retrieval, and cross-channel messaging (primarily WhatsApp) under a single permissions model.
The AI is a component. Subport is the chassis, the wiring, and the dashboard.
The Framing Problem
It's tempting to look at Subport and think: "This is just an AI agent with some tools. I can build that with Claude Code and a handful of MCPs."
And in a narrow sense, you're right. Claude Code + MCPs can:
- Read from databases
- Call APIs
- Search documents
- Execute code
- Maintain context within a session
But that combination solves a developer's problem — "let me automate this task." Subport solves an organization's problem — "let me run a 24/7 support operation across hundreds of customers, dozens of agents, and multiple channels without everything falling apart."
The gap is not intelligence. The gap is infrastructure, state, and orchestration at scale.
What Subport Actually Builds (That Claude Code + MCPs Doesn't)
1. Persistent Multi-Tenant Runtime
Claude Code runs in a session. Subport runs in perpetuity.
Every conversation, ticket, customer record, and knowledge base article lives in a workspace → project → tenant hierarchy with strict data isolation. A customer messaging you on WhatsApp today is the same customer who emailed last week — their history, preferences, and open tickets are resolved automatically. Claude Code has no concept of "customer identity across time." Subport's entire data model is built around it.
The value: You don't rebuild auth, isolation, and identity for every new use case. It's already there.
2. Stateful Conversation Orchestration (Days, Not Minutes)
A Claude Code session dies when you close the terminal. A Subport conversation can span hours, days, or weeks.
The system maintains:
- Conversation context across arbitrary time gaps
- Message threading with inbound/outbound tracking
- Entity resolution (tickets, customers, assignments) injected into every AI call
- Cross-channel continuity (start on WhatsApp, continue via Slack, escalate to a human in the dashboard)
This isn't "memory" in the LLM sense. It's a conversation state machine with explicit handoff points, timeout handling, and resume capability.
The value: Real support doesn't happen in a single chat window. It happens across shifts, time zones, and interruptions.
3. Human-in-the-Loop as a First-Class System
Subport doesn't treat human escalation as a fallback. It treats it as a governed protocol.
The HITL system includes:
- Risk-level classification (low / medium / high) for agent actions
- Structured confirmation requests with timeout and expiry
- Permission-aware approval chains (a junior agent's AI can't approve refunds; a manager's can)
- Audit trails for every human decision injected back into the conversation
This isn't "the AI got stuck, let me take over." It's "the AI knows when it's about to do something sensitive, pauses, notifies the right human through the right channel, and resumes only after documented approval."
The value: In regulated or high-stakes environments, you need structured human oversight, not ad-hoc intervention.
4. Channel Abstraction Layer
Claude Code reads files and runs shell commands. Subport sends and receives messages on WhatsApp Business API, Slack, and other channels — handling:
- Message format translation (WhatsApp buttons, lists, template messages)
- Media upload/download with cloud storage
- Delivery status tracking and retry logic
- Webhook signature validation and rate limiting
- Phone number mapping to customer identities
The WhatsApp Business API alone is a beast: conversation-based pricing, template approval, 24-hour session windows, interactive message types. Subport abstracts all of that into a unified ChannelAdapter interface.
The value: Your business logic doesn't change when WhatsApp updates its API. The adapter does.
5. Knowledge Base RAG (Production-Grade)
Yes, you can connect Claude Code to a vector DB via MCP. Subport's knowledge base goes further:
- Multi-format ingestion: PDF, DOCX, Markdown with intelligent chunking
- pgvector similarity search with configurable relevance thresholds
- Per-project KB scoping (customer A can't see customer B's internal docs)
- Tool-friendly identifiers — the AI references KBs by slug (
search_knowledge_base(knowledgeBase: "returns-policy")) - Context assembly that injects KB results into system prompts before every AI call
This isn't "I uploaded a PDF and asked questions." It's "every conversation automatically queries the right knowledge base, filters by permissions, and grounds responses in documented truth."
The value: RAG that respects tenant boundaries and scales to thousands of documents without manual indexing.
6. SLA Orchestration and Ticket Lifecycle
Subport has a multi-policy SLA engine that most "AI support demos" never even consider:
- Priority-based SLA selection
- Business hours + timezone awareness
- Breach detection and automatic escalation
- Auto-close with warning timers
- Reopen handling and re-breach logic
Tickets aren't just database rows. They're stateful entities with assignment rules, custom field definitions, category schemas, and lifecycle workflows. The AI interacts with this system through structured tools — it doesn't just "know" about tickets, it operates on them through a typed API.
The value: An SLA breach at 2 AM doesn't wait for a human to notice. The system escalates, reassigns, and notifies automatically.
7. Durable Workflow Engine
Built on Vercel Workflows, Subport's flow engine supports:
- Checkpoint-based resumption for long-running jobs
- Parallel, loop, switch, and delay nodes
- Immutable published snapshots (draft → publish → execute)
- Trigger-based execution (webhooks, schedules, database events)
- Visual builder for non-technical users
This is not "the AI decides what to do next." It's "a human designs a business process, publishes it, and the runtime guarantees execution even across server restarts."
The value: Business processes survive crashes, retries, and delays. The runtime is durable, not just the model.
8. Permission-Aware AI Execution
Here's where things get subtle and important. In Subport:
- Users get permissions through Groups only — no direct user→permission grants
- Agent capabilities = Agent Capabilities ∩ User Permissions
- Tool execution carries the invoker's resolved permissions into the action
If a customer-group user triggers an AI agent, the agent literally cannot see fields or perform actions that user isn't allowed to access — the permission set is threaded through the entire execution context. There's no "AI bypass."
Claude Code + MCPs has no such concept. If you give the AI a database connection, it has the connection's permissions. There's no intersection model, no group-based RBAC, no field-level visibility.
The value: You can safely expose AI tools to customers without fearing privilege escalation. The AI is bound by the same authorization lattice as the human user.
The Honest Assessment: Where Is the Real Value?
Real Additional Value: Yes, But It's Infrastructure
Subport's moat is not "we have an AI that can answer questions." Its moat is:
- The data model — Tickets, customers, conversations, SLAs, projects, workspaces. This is years of domain modeling, not a weekend project.
- The channel integrations — WhatsApp Business API complexity alone justifies a platform.
- The permission system — Group-based RBAC with agent intersection is genuinely hard to bolt on later.
- The durable runtime — Workflow execution that survives restarts and retries.
- The UI for non-technical operators — Support managers need dashboards, not terminal sessions.
Where the Value Is Not
- Raw AI capability: Subport uses off-the-shelf models (GPT-4, Claude, Gemini via AI SDK). It doesn't train its own. The model is interchangeable.
- Novel prompting techniques: The system prompts and RAG assembly are solid engineering, not research breakthroughs.
- Single-session automation: If your task is "analyze this one file," Claude Code wins.
The Real Question
The question isn't "Can Claude Code + MCPs do what Subport does?" The question is "How many engineering-months does it take to turn Claude Code + MCPs into Subport?"
The answer, based on the codebase: a lot. This is a production SaaS with:
- Multi-tenant auth and authorization
- Database schema organized by domain
- Integration test suites with VCR recording
- E2E Playwright suites
- Structured logging and monitoring
- Vector search infrastructure
- WhatsApp gateway service
- Workflow runtime with checkpointing
You don't MCP your way to that.
The Long-Term Risk (And Why It Doesn't Change the Answer)
As MCP ecosystems expand and models get agentic native capabilities, some of Subport's "glue code" will become thinner. A future Claude might handle multi-day sessions, built-in RBAC, or native WhatsApp integration.
But even then, Subport retains value in:
- The ticket/SLA domain model — AI models don't know what an SLA breach policy should look like. That's business logic.
- The human operator UI — Someone still needs to see the queue, assign tickets, and review escalations.
- Compliance and audit trails — Enterprise buyers need proof, not just functionality.
- Reliability guarantees — A local Claude session doesn't come with an SLA. Subport does.
Bottom Line
Subport is not "Claude Code for customer support." It's "the infrastructure you need so that Claude (or GPT-4, or Gemini) can safely and reliably operate in a multi-tenant, multi-channel, SLA-bound production environment."
The AI is the engine. Subport is the car — complete with seatbelts, brakes, navigation, and a dashboard. You can build a car around an engine, but the engine alone won't get you to work.
If your problem is "I need to automate a task," use Claude Code + MCPs.
If your problem is "I need to run a 24/7 AI-powered support operation for hundreds of customers without losing data, violating permissions, or missing SLA deadlines," you need the infrastructure. That's Subport.