Omniagent

One agent. Every channel. Every job.

Omniagent is an AI-agent product for real businesses. It deploys agents that answer customers and run the back office over WhatsApp — understanding text, voice notes, images, PDFs and spreadsheets, answering from the business's own knowledge (RAG), remembering the conversation, and taking real actions in real tools with human approval on anything that moves money.

Built WhatsApp-first; the same agent core is channel-agnostic by design. The engine is a set of importable n8n workflows — a customer runs it on their own n8n, OpenAI key and MongoDB Atlas cluster.

---

What's in this repo

Path	What it is
index.html	The Omniagent marketing & sales site — self-contained, no build step. Open it in a browser or serve the folder. This is what a prospective business sees.
omniagent-engine/	The agent engine — importable n8n workflows that run the real product (multimodal RAG pipeline + memory + gated tool actions). See its README.
docs/	Go-to-market collateral: the sales one-pager and the deployment runbook.
mcp-server/	A Model Context Protocol server that lets an AI assistant inspect the agent catalog, personas, tools and docs in this repo. See its README.

The agent catalog

Every agent shares the identical multimodal front-end (WhatsApp trigger → type routing → voice/image/document handling → unified prompt → agent → reply). Only the persona and the tools change — so a new agent is configuration, not a rebuild.

Agent	Workflow	What it does
Concierge	omniagent-engine/workflow.json	Always-on first responder — grounded answers over every format, clean escalation.
Support Agent	omniagent-engine/workflow-customer-support.json	Front-line support for any business + ticket/escalation tool.
Bookkeeper Agent	omniagent-engine/workflow-quickbooks-specialist.json	Live QuickBooks Online reads + CONFIRM-gated invoice/payment writes.
Reservations Agent	omniagent-engine/workflow-reservations.json	Real availability checks + CONFIRM-gated booking / reschedule / cancel.
Sales Qualifier	omniagent-engine/workflow-sales-qualifier.json	Greets & qualifies inbound leads, books demos, pushes scored leads to the CRM.

The specialists are generated from the base workflow — edit omniagent-engine/workflow.json, run node build-variants.mjs, and every agent inherits the change.

Architecture

flowchart LR
    WA["WhatsApp Cloud API\n(text · voice · image · doc)"] --> RT["Route by type"]
    RT -->|voice| TR["Transcribe (OpenAI)"]
    RT -->|image| VI["Vision analysis"]
    RT -->|doc| EX["Extract PDF / XLS / XLSX"]
    RT -->|text| P["Unified prompt"]
    TR --> P
    VI --> P
    EX --> P
    P --> AG["Agent (persona + tools)"]
    LM["OpenAI chat model"] --- AG
    MEM["Conversation memory\n(100-turn window)"] --- AG
    RAG["MongoDB Atlas\nVector Search (RAG)"] --- AG
    TOOLS["Domain tools\n(QuickBooks · bookings · CRM · tickets)\nwrites CONFIRM-gated"] --- AG
    AG --> OUT["WhatsApp reply"]

Loading

Stack: n8n + OpenAI + WhatsApp Cloud API + MongoDB Atlas Vector Search. The ingest-knowledge-base.json companion workflow loads a business's documents into the vector store.

Run the site locally

No build step — it's a single self-contained file (fonts from Google Fonts, nothing else).

python3 -m http.server 4600   # then open http://localhost:4600

index.html is what GitHub Pages serves (deployed automatically on every push to main by .github/workflows/pages.yml).

MCP server (AI-native repo)

The repo ships a stdio MCP server so an AI assistant can query it as structured data instead of raw files:

Tool	Answers
list_agents	"What agents exist, on which model, with which tools?"
get_agent	"What exactly is the Bookkeeper's persona and which writes are gated?"
diff_agent_vs_base	"How does a specialist differ from the base Concierge?"
get_site_structure	"What does the sales site actually say, section by section?"
get_doc / search_repo	Docs verbatim, and search across everything.

cd mcp-server && npm install && cd ..
claude mcp add choreless -- node mcp-server/server.mjs

A root .mcp.json registers it for clients that support project-scoped MCP config. Details in mcp-server/README.md.

Why businesses buy it

• Grounded, not guessing — answers are retrieved from the customer's own documents and cited inline; if it's not in the knowledge base, the agent says so and escalates.
• Human-gated actions — anything irreversible is held behind a typed CONFIRM protocol baked into the agent's system prompt.
• Their data, their tenancy — runs on the customer's own n8n, OpenAI key and MongoDB Atlas cluster. No middleman, no training on their conversations.
• Live in a weekend — one importable workflow, ~20 minutes of credential wiring.

See docs/SALES-ONEPAGER.md for the pitch and docs/DEPLOYMENT.md for the go-live runbook.

Status & history

This is a product demo / go-to-market prototype: the workflows are real and importable, but you bring your own n8n instance, WhatsApp Business number, OpenAI key and MongoDB Atlas cluster — the deployment runbook walks through all of it. The repository is named choreless because it previously hosted the Choreless chore-outsourcing concept; it was repurposed for Omniagent in July 2026 (the history is in the git log).

License

MIT © 2026 GreenAI Solutions.