For many users seeking artificial intelligence assistance with travel planning, destination research, and itinerary building, the standard cloud-based ecosystem often presents a growing financial and privacy hurdle. Navigating subscriptions, recurring monthly fees that scale with usage, and sending personal queries to external cloud providers drives a constant search for alternatives. This exact technical challenge recently inspired software engineer and technical writer Shittu Olumide to develop a fully functional, local multi-agent travel research team running entirely on consumer hardware without subscriptions or external cloud dependency.
The resulting project, known conceptually as Wanderlist, relies on the integration of two powerful open-source tools: Hermes Agent and Ollama. By pairing an advanced open-source AI agent framework with a local model runner, developers and enthusiasts can build customized, multi-agent teams capable of executing real-world tasks such as web browsing, file management, and tool execution directly on a personal computer.
To understand how the Wanderlist project functions, it is essential to examine the two primary technologies underpinning its architecture.
Introducing Hermes Agent
Hermes Agent is an open-source AI agent developed by Nous Research and distributed under the permissive MIT license. Available as both a native desktop application for macOS, Windows, and Linux, and a terminal-first command line interface, Hermes provides genuine agentic capabilities that distinguish it from conventional conversational chatbots.
Unlike standard chat interfaces that merely respond to text prompts with pre-trained knowledge, Hermes can read and edit local files, execute terminal commands, and browse the live web. This combination of capabilities makes it well-suited for dynamic tasks like travel planning, where an agent must retrieve current, real-time information rather than relying on static training data.
Several key features enable sophisticated project architectures within Hermes. Multi-agent orchestration allows a coordinator agent to delegate specific sub-tasks to specialist agents and synthesize their findings into a cohesive output. Persistent memory enables the agent to retain context across separate sessions, recalling stated budgets, travel styles, or past itineraries rather than resetting with every new prompt. Furthermore, Model Context Protocol (MCP) tool support connects Hermes to external servers, including isolated web-search containers, while custom tool support allows developers to integrate bespoke Python functions directly into the agent’s workflow.
Understanding Ollama and Local Models
Providing the cognitive foundation for Hermes is Ollama, an open-source tool designed to download, manage, and serve large language models locally on a user’s computer. Ollama exposes models through a local application programming interface that mirrors standard cloud provider endpoints, substituting external URLs with a local address pointing to localhost. Because Ollama’s API is compatible with OpenAI integration paths, Hermes can interact with locally running models seamlessly.
Running local models means utilizing open-weight parameters published by researchers, operating entirely on a user’s local central or graphics processing unit. Nothing leaves the machine except specific web searches initiated by the agent, eliminating usage-based billing in favor of standard electricity costs. While consumer hardware imposes practical limits on raw processing speed and capacity compared to massive frontier cloud models, the trade-off is often ideal for structured tasks involving research, planning, and tool utilization.
Designing the Travel Team Architecture
The Wanderlist project implements a hierarchical agent structure led by a root coordinator supported by three distinct specialist agents. The root coordinator receives initial user prompts regarding a desired trip and delegates specific responsibilities outward before compiling the final results into a comprehensive briefing.
The first specialist, designated as the destination-scout, is tasked with researching target locations, notable attractions, and current entry requirements. To ensure accuracy, instructions mandate that the scout perform live web searches for visa and entry rules rather than relying on static model memory, mitigating the risk of outdated information impacting travel plans.
The second specialist, the stay-finder, investigates well-regarded neighborhoods, accommodation styles, and typical price ranges using recent travel sources, operating under strict constraints against inventing specific unverified listings.
The third specialist, the flight-finder, generates functional flight search links and evaluates current approximate fare contexts. This agent pairs standard web-search capabilities with a custom Python tool designed to construct verified Uniform Resource Locators for major flight search platforms.
Implementing Custom Tooling and Persistent Memory
To bridge the gap between AI generation and practical utility, the project incorporates a custom Python utility that builds genuine search URLs for Google Flights. Because real-time, free flight-booking APIs are largely unavailable, generating natural-language query parameters via URL encoding provides a reliable, keyless alternative. The utility validates user inputs, ensuring that origin and destination fields are populated and that travel dates conform strictly to standard year-month-day formats, raising immediate exceptions for malformed data rather than producing broken links.
Additionally, the coordinator agent utilizes a persistent database for trip preferences. By storing user parameters such as budget constraints, accommodation preferences, and historical interests across sessions, the system adapts over time, reducing repetitive prompting and improving personalization with subsequent use.
Executing a travel planning session involves initiating the Hermes run command with the configured team parameters. When presented with a prompt detailing a destination, travel window, origin, and user preferences, the coordinator distributes tasks among the destination scout, stay finder, and flight finder. The resulting output delivers a structured briefing complete with sourced attraction data, accommodation guidance, and a clickable search link, demonstrating the viability of agentic workflows executed entirely on local hardware.
