使用说明

Agentic AI with MongoDB - A Technical Showcase

Demonstrating a next-generation AI architecture with self-organizing agents that discover services, maintain memory, and orchestrate complex workflows using MCP, MongoDB Vector Search, and Voyage AI.

How to install and run the demo

Setup is straightforward. Clone the repository, and follow these steps:

Set environment variables for aaS components

export OPENAI_API_KEY="<your openai api token>"
export MONGODB_URI="<your mdb connection string>"
export VOYAGE_API_KEY="<your voyage api token>"   # used by the restaurant_guide MCP service

Install Python 3.13

brew install python@3.13

Set the path in your .zshrc

export PATH="$(brew --prefix)/opt/python@3.13/libexec/bin:$PATH"

Create and activate a Python virtual environment

python -m venv <dir>
source <dir>/bin/activate

Install Python packages

cd agentic-mcp-demo
pip install -r requirements.txt

Run the shell and interact

python main.py

Look behind the scenes

The external live feed is optional. Point NOTIFY_BROADCAST_URL / NOTIFY_RECEIVE_URL at your own notify relay (unset by default → the external POST is skipped; the in-browser Agent Log still works). Then open another shell and run:

curl -sN "$NOTIFY_RECEIVE_URL" | sed -n 's/^data: //p'

Watch the portfolio live

The portfolio MCP service ships with a web dashboard backed by MongoDB Change Streams + WebSockets. Run it in a separate shell:

python web/portfolio_dashboard.py

Then open http://localhost:8050 — every add/update/delete from the agent flows in real time without polling.

Have fun! ;-)

Why MongoDB

360-degree data foundation / digital twin

A multi-agent, multi-MCP architecture is only as strong as its underlying data foundation. This foundation must be accessible in a highly performant, resilient manner and as comprehensive as possible. MongoDB's Document Model excels at building such a base, enabling rapid iteration during innovation and experimentation.

Voyage AI

Voyage AI's powerful embedding models drive MCP service discovery. Moreover, select MCP servers employ them to enable semantic search across multi-modal content, facilitating precise retrieval, processing, and transactional operations.

Short/Long-Term Memory and User Preferences

MongoDB serves as the cohesive layer linking MCP server invocations, overseeing agentic memory services. It manages both enduring data like user preferences ("I am vegetarian," "I often travel overseas") and transient situational details ("today I want Indian food," "looking for a new iPhone").

TTL Indexes

Short-term memory undergoes automatic garbage collection after a set period, powered by MongoDB's TTL indexing. This eliminates the need for external scheduled scripts.

Geospatial Indexes

Some MCP servers integrate geospatial search into their operations. In the Telco/OSS demo flow, for example, one server maps a customer's current location to nearby cell towers, enabling discovery of proper functioning, and conducting service assurance.

Implemented demo flows (so far, more to come)

Telco/BSS billing / plan management

You interact with the following questions/commands:

Show my current invoice
Why so expensive
Options to reduce costs
Book addon world roaming plus
Confirm

Intent-Based Networking (IBN) — retail chain onboarding

The flagship technical demo. A retail chain customer expresses a business intent in plain language; the system parses it, checks feasibility against live inventory, generates a plan, activates it, and continuously defends the promise. When a violation fires, a single MongoDB Atlas aggregation pipeline combines semantic similarity + geospatial pre-filter + time window + structured equality to surface the closest matching past incident and its proven runbook — the "institutional memory" moment that makes MongoDB Atlas hard to replace with a Postgres + pgvector + PostGIS + TimescaleDB cocktail.

Setup (one time):

python seed/ibn_seed.py

This loads the fixtures (5 sites with geo, ~20 vector-indexed knowledge chunks, 4 active intents) and prints the JSON config for the Atlas Vector Search index you need to create on ibn_knowledge_chunks in the Atlas UI. Until that index is built, the diagnose tool will fail.

Run the live dashboard (separate shell):

python web/ibn_dashboard.py

Then open http://localhost:8060 for engineer view, or http://localhost:8060/?mode=exec for executive view. Same backbone, two rendering styles.

5-minute demo conversation:

You: I'm opening a new Alpenmarkt store at Marienplatz Munich.
     POS priority, guest WiFi strictly separated, camera uplink,
     online by 18:00, max 40ms POS latency, 99.95% availability.
You: Is this feasible?
You: Activate it.
You: Inject morning rush.
You: What happened?           ← this fires the WOW vector query
You: Apply the runbook.

10-minute extended flow adds: show all active intents, resource inspection at the new site, geospatial failover lookup, post-violation fleet breadth check, runbook preview, and update the template so future stores inherit the fix — closing the institutional-memory loop.

Telco/OSS network assurance / customer complaint management

You interact with the following questions/commands:

Check network status for customer +49 176 12345678
Check tower status. And just in case, create ticket, and apply a credit to this premium customer

Use Cases

Not all are part of the demo yet - but should, over time. Please contribute MCP servers and demo flows!

Telco/BSS

Autonomous Quote‑to‑Order Orchestration

Agentic systems interpret business intent from upstream channels, validate eligibility, configure multi‑domain services, and orchestrate end‑to‑end order execution. They coordinate across CRM, CPQ, fulfillment, and network layers—resolving exceptions in real time to minimize latency in the commercial‑to‑operational flow.

Intelligent Revenue Assurance & Leakage Prevention

Agents continuously monitor mediation, rating, billing, and settlement streams to detect usage anomalies and reconciliation gaps. They apply rule‑ and model‑driven diagnostics, initiating automated corrective actions and feeding insights back into assurance and policy engines.

Adaptive Product & Pricing Optimization

Autonomous policy agents analyze demand and margin signals across offers and usage segments. Through API‑level interaction with product catalogs and charging systems, they dynamically propose pricing adjustments, bundle configurations, or discount rules aligned with operational and financial targets.

Proactive Lifecycle Orchestration

Agents coordinate service lifecycle events across provisioning, assurance, and billing domains. Using predictive indicators from telemetry and transaction flows, they preempt degradation, trigger workflow adjustments, and ensure closed‑loop fulfillment and assurance across the BSS/OSS stack.

Partner & Ecosystem Automation

In multi‑party and B2B2X environments, agents manage partner onboarding, contract compliance, usage settlement, and dispute resolution. They operate as autonomous orchestration nodes—negotiating API contracts, validating transactions, and maintaining distributed state synchronization across ecosystem participants.

Telco/OSS

Network Fault & Performance Management

Autonomous agents correlate FCAPS streams (SNMP/TL1/gNMI) in real-time across topology-aware memory. They perform root-cause inference via ML-based anomaly detection, prioritize alarms by impact, and trigger self-healing—from NE reconfiguration to dispatch orchestration.

Service Activation & Assurance Orchestration

Intent-based agents decompose service orders into multi-domain workflows (L1-L7). Continuous telemetry validates post-provisioning KPIs, predicts degradation, and executes autonomous rollbacks or remediation—leveraging TMF814/815 APIs.

Capacity & Resource Planning Optimization

Agents consume live PM data and demand forecasts to simulate capacity scenarios via graph neural networks. They negotiate with inventory/procurement systems, trigger equipment orders, and optimize fiber/circuit routing just-in-time.

Network Inventory & Reconciliation

Discovery agents maintain distributed golden records through continuous NE sweeps and event-driven updates. Reconciliation detects drift in real-time, infers causes via change-tracking, and executes bulk updates or reprovisioning with escalation.

Interconnect & Wholesale Settlement

Agents monitor peering/wholesale CDR streams live, validate against smart contracts, and compute charges on-the-fly. They resolve disputes autonomously and generate TMF-compliant settlement files through continuous contract synchronization.