AI infrastructure · WhatsApp · LATAM

Persistent memory for AI agents on WhatsApp.

The layer your agent needs to remember customers, extract orders, and operate businesses in Mexico and LATAM. Not a messaging API.

Start free See capabilities Talk to sales
  • MCP-native
  • Vector search
  • SAT · CFDI · SPEI · OXXO
Live infrastructure
20MCP tools
6categories
100%audited tools
2transports (HTTP/SSE · stdio)
MCP capabilities

Six categories. One API.

Every capability exposes MCP tools ready to be called by your agent. All audited, scoped, and idempotent.

WhatsApp Messaging

4 tools

Send text, Meta-approved templates, and media. Opt-out enforced, 24h-window applied, queued via BullMQ with retry. Idempotent.

  • whatsapp_send_message
  • whatsapp_send_template
  • whatsapp_send_media
  • whatsapp_template_submit

Customer & Catalog

3 tools

Look up contacts, search the product catalog (semantic with embeddings or LIKE fallback) and inventory by SKU. Real data from Prisma.

  • customer_lookup
  • product_search
  • inventory_check

Orders

2 tools

Create draft orders (never auto-confirmed). Confirmation requires explicit customer or operator evidence. Idempotent.

  • order_create_draft
  • order_confirm

Payments & Invoicing

4 tools

Stripe (incl. OXXO Pay via Stripe MX), Mercado Pago, SPEI with unique reference. CFDI 4.0 via Facturapi. HMAC-verified webhooks.

  • payment_link_create
  • spei_instructions_generate
  • payment_status_check
  • invoice_create

Conversations & Memory

5 tools

Heuristic structured conversation digest, full-text search across messages, and per-contact key-value memory with TTL.

  • conversation_summarize
  • conversation_search
  • memory_set
  • memory_get
  • memory_list

Handoff & Compliance

2 tools

Escalate conversation to a human with structured summary (disables the agent). Record consent opt-in/opt-out with auditable evidence (LFPDPPP).

  • handoff_to_human
  • consent_update

Full catalog in mcp-manifest.json and /.well-known/mcp.json.

How memory works

Before vs. after TramsTec.

The difference is not marginal. Without memory, your agent starts from zero every message. With TramsTec, it remembers and reasons across the entire history.

Without memory

  • "Hi, how can I help?" to returning customers
  • Orders buried in chat, never extracted
  • Zero context from past quotes
  • Impossible to search "what did I offer Pedro?"
  • Agent repeats questions it already asked
  • More tokens (you re-feed context every time)

With TramsTec

  • "Hi María, want to repeat the 2kg tortillas?"
  • Orders auto-extracted and structured
  • Quotes and proposals indexed by intent
  • Semantic vector search in <50ms
  • Real continuity across conversations
  • Your agent only asks what it needs
End-to-end memory flow Customer sends a WhatsApp message → TramsTec indexes it in pgvector and extracts intent/entities → your agent calls getCustomerContext and receives memory + context → replies with continuity. 1 CUSTOMER WhatsApp · 6:42pm "I want the usual please" ✓✓ 2 TRAMSTEC LIVE indexMessage() ✓ 12ms extractIntent() ✓ 24ms searchMemory() ✓ 38ms pgvector: hit · sim 0.94 3 YOUR AGENT receives context { name: "María García", lastOrder: "2kg", intent: "repeat", sentiment: "positive" } 4 REPLY "Hi María! 2kg as last time?" 6:42pm ✓ All persists for the next conversation · automatic pgvector indexing · multi-tenant isolated per org
LATAM workflows

Real use cases, not demos.

Each workflow chains several MCP tools. Your agent orchestrates them. TramsTec handles state.

🛒

Retail order with payment

  1. Customer asks for products via WhatsApp
  2. Agent looks up customer + catalog + inventory
  3. Creates draft order and generates payment link (Stripe / Mercado Pago / SPEI)
  4. PSP webhook confirms payment → agent confirms order and notifies the customer

customer_lookup · product_search · inventory_check · order_create_draft · payment_link_create · order_confirm · whatsapp_send_message

💰

Stage-aware collections

  1. Look up customer + pending payment status
  2. Generate SPEI instructions with a unique reference
  3. Send reminder using a Meta-approved template
  4. If no response, escalate to human with context

customer_lookup · payment_status_check · spei_instructions_generate · whatsapp_send_template · handoff_to_human

🧾

Mexico CFDI invoice

  1. Customer requests invoice in chat
  2. Agent looks up customer + referenced payment status
  3. Generates CFDI 4.0 via Facturapi (org's RFC)
  4. Sends signed PDF over WhatsApp media

customer_lookup · payment_status_check · invoice_create · whatsapp_send_media

🛡

Auto opt-out (LFPDPPP)

  1. Inbound message detects "BAJA" / "STOP" / "UNSUBSCRIBE"
  2. Agent records a consent event with auditable evidence
  3. Any future send to this contact is runtime-blocked
  4. Audit log preserves the evidence indefinitely

consent_update

💬

Sales follow-up (Roadmap)

  1. Detect unanswered quotes
  2. Recover context from previous conversations
  3. Re-engage with personalized message
  4. Escalate to human if lead cools

Requires extract_purchase_intent + get_pending_follow_ups (Q3 2026)

🎯

Lead qualification (Roadmap)

  1. New lead arrives via WhatsApp
  2. Agent extracts vertical, budget, urgency
  3. Assigns score and tag
  4. Routes to right human or nurture sequence

Requires extract_lead_insights + score_customer (Q3 2026)

AI augmentation

Not just plumbing. Each tool, smarter.

You bring the main LLM (Claude, OpenAI Agents, LangChain). TramsTec embeds specialized micro-models inside specific tools — embeddings, transcription, voice synthesis, extraction — making each one measurably smarter than a generic WhatsApp wrapper. Marginal cost per customer: less than $5 USD/month.

Live

🔎 Semantic catalog search

Customer types "something sweet for dessert" → finds Tres leches, Flan napolitano, Cheesecake even though those products don't contain "sweet" or "dessert" in their names.

OpenAI text-embedding-3-small + pgvector · tool: product_search

Q3 2026 · Flagship LATAM

🎤 Voice notes IN — auto-transcribed

60%+ of WhatsApp messages in Mexico are voice notes. Your agent gets them automatically transcribed via Whisper, with metadata original_type: "voice".

OpenAI Whisper · ~$0.001 per voice note · webhook processes before passing to agent

Q3 2026 · Flagship LATAM

🔊 Voice notes OUT — agent replies in voice

Your agent responds with natural voice notes in Mexican Spanish (or other locales). Customer feels they're being helped by a person, not a bot. Matches the channel norm.

OpenAI TTS / ElevenLabs · new tool: whatsapp_send_voice

Q3 2026

🎯 Extract purchase intent

Parses free-form WhatsApp text into structure: {category, priceRange, urgency, confidence}. Your agent needs no prompt engineering — just call the tool.

tools: extract_purchase_intent, extract_lead_insights, extract_orders_from_conversation

Q3 2026

📊 Customer relationship summary

Auto-generated executive summary per customer: LTV, frequency, favorites, churn risk, upsell signals. Real CRM from raw conversations.

tool: customer_relationship_summary

Q4 2026

📝 LLM conversation summarization

Current conversation_summarize uses heuristics. The LLM version generates executive digests for human handoff, reports, and dashboards.

upgrade of existing tool · OpenAI / Claude / Mistral configurable

Philosophy: You bring the big LLM for conversation. We embed micro-models inside specific tools — invisible to the agent but measurably smarter than a wrapper. Any customer activating these features pays ~$5 USD/mo extra in OpenAI/ElevenLabs costs, on top of the plan price. To accelerate any roadmap item with a paid pilot, contact sales@tramstec.com.

LATAM countries

Built for LATAM reality.

85% of Mexican SMBs sell on WhatsApp. Global infra doesn't understand that. We do.

🇲🇽 Mexico 🇦🇷 Argentina 🇨🇴 Colombia 🇨🇱 Chile 🇵🇪 Peru 🇧🇷 Brazil · soon 🇻🇪 Venezuela · roadmap
Your stack

Plug-and-play with any agent.

Native MCP + REST + SSE. SDKs in TypeScript and Python.

Build vs. Buy

How many engineer-months before production?

It isn't "can I build it?" — it's "how many sprints will you burn before your agent runs in earnest?".

CapabilityIn-houseTramsTec
WhatsApp Cloud API + reconnect3-6 wkbuilt-in
Semantic memory (vector)1-2 mobuilt-in
Auto-generated CRM1-3 mobuilt-in
Order / lead extraction1-2 motool
Semantic search3-6 wktool
Commercial continuity1-2 motool
24h window + templates2-3 wkauto
Idempotency + audit3-5 wkbuilt-in
Multi-tenant + API keys1-2 mobuilt-in
CFDI 4.0 + RFC + SAT2-4 wkbuilt-in
SPEI + OXXO recon3-6 wkbuilt-in
Observability + on-callongoingwe run it
Total to prod6-12 moone afternoon
Infrastructure

Infrastructure-grade, not toy project.

Built for agents in production, not for demos. We run it the way you would — if you had the team.

HMACMeta + Stripe webhooks
Auditimmutable tool_calls
24hidempotency window
Multitenant org-isolation
ScopedAPI keys + per-tool scope
BullMQqueue + exp retry
SOC2on roadmap (2027)
PostgresJSON-portable
Autonomous discovery

Designed so agents can find us on their own.

Any AI agent or LLM crawler can understand this service without human intervention. These are the public artifacts we expose:

📡 Discovery endpoints

Machine-readable manifests + structured metadata on every page.

Pricing

Pricing designed for AI agents

Pay for operational infrastructure, not for seats. Start free, deploy to production, and scale to millions of WhatsApp conversations.

Developer

Free

Start building

  • Sandbox environment
  • MCP-ready documentation
  • Test WhatsApp workflows
  • Build and validate your agent
  • Ideal for developers experimenting with WhatsApp-native AI
Production

$99 /month

Launch real AI agents on WhatsApp

  • Production-ready WhatsApp infrastructure
  • Context memory for business conversations
  • Reliable message orchestration
  • Agent-to-human handoff support
  • Analytics for live conversations
Platform

$999 /month

Scale thousands of conversations

  • Multi-tenant infrastructure
  • White-label ready
  • Advanced observability
  • API access for platforms and agencies
  • Priority support for mission-critical agents
Enterprise

Custom

Custom infrastructure for large-scale AI operations

  • Dedicated deployment options
  • Custom SLAs
  • Security and compliance review
  • Advanced integrations
  • Enterprise onboarding and support
Contact Us

Starting prices listed in USD. Usage-based billing on tool calls, memory operations and active conversations. WhatsApp/Meta fees and payment gateway commissions pass through at cost.

Get started

Your agent on memory in minutes.

Get on the waitlist and we'll email early access: API key, MCP endpoint, starter snippet for your stack.

Control & portability

Your data. Your agent. Your rules.

Zero lock-in. Every decision audited. Export or revoke any time.

Scoped API keys

One per agent, client, or environment. Revoke in 1 click.

Immutable audit

Every tool call with timestamp, args, result.

Portability

Export all your memory in JSON or Parquet.

Build agents with memory. In LATAM.

Start free. No card.

Reserve my spot See capabilities Sales