Why Captain Went Open Source
Captain at getmtp.com went open source because travel planning needs inspectable orchestration, not just a confident model. In 2025, frameworks that deliver combine LLM reasoning with deterministic workflows, persistent memory, and real-time communications. Burr offers a full-stack agent framework for stateful graphs; Novyx adds memory APIs for rollback, replay, and semantic search. LiveKit Agents brings real-time outbound voice AI to cloud infrastructure, crucial when agents call hotels or airlines. American Express's Intent Intelligence Framework signals that trusted agentic commerce requires explicit intent, permissions, and audit trails.
Also worth reading: Can an AI travel packing assistant actually pack smarter for your next trip? · What Actually Happens During Official Travel Document Checks in 2026? · What Are the Best AI Travel Planning Tools in 2026, and Which Ones Are Actually Worth Using?
The practical winner is a stack, not a single framework. Use Burr for durable execution, Novyx for memory and replay, LiveKit for voice, and a commerce-aware intent layer for trust. Captain's open-source turn lets developers optimize and launch a travel-planning AI application in minutes while keeping costs, latency, and failures visible. The frameworks that deliver in 2025 treat LLMs as one component, connect to real inventory and booking APIs, and let humans intervene when plans change.
Beyond LLMs: Memory and State
The frameworks that actually deliver in 2025 are the ones treating memory as infrastructure rather than an afterthought. Novyx's memory API, with rollback, replay, and semantic search, points to what travel agents genuinely need: the ability to remember a traveler's aisle preference from three trips ago, recover gracefully when a booking fails mid-transaction, and replay a session to debug why the agent suggested a red-eye. Burr's full-stack approach and LiveKit's real-time voice agents on Oracle Cloud show the same lesson from different angles. Captain and similar Show HN travel agents succeed precisely because they wrap the LLM in durable state, not because their prompts are cleverer.
American Express's Intent Intelligence Framework signals where enterprise trust is heading: verifiable agent actions, not fluent conversation. For travel specifically, the stakes are high because itineraries span days, vendors, and currencies, and a hallucinated confirmation number is a real financial loss. The winning stack pairs a capable model with persistent memory, deterministic tool calls, and observability. Optimizing a travel-planning app in minutes is now trivial; making it reliable across a two-week trip is the actual product.
Agent Frameworks Compared for Travel
Which AI Travel Agent Frameworks Actually Deliver in 2025? The honest answer is that most general-purpose agent frameworks still stumble on travel's core demands: multi-step itineraries, live pricing, cancellation policies, and the messy reality of rebooking a red-eye after a missed connection. Captain, the Show HN travel agent, leaned on LLMs for conversation but wrapped them in deterministic booking logic, which is why it worked where pure prompt-chaining demos failed. Burr (OS) offers a full-stack state machine approach that handles branching plans well, though travel-specific tooling remains thin. Novyx's memory API addresses a real gap—rollback and replay matter when an agent books the wrong hotel, and semantic search over past trips beats naive context stuffing.
The frameworks that actually deliver in 2025 share a pattern: they treat the LLM as a reasoning layer, not the whole system. American Express's Intent Intelligence framework signals where enterprise trust is heading, while LiveKit's real-time voice agents on Oracle Cloud show that outbound confirmation calls are now viable. For travel specifically, the winners are hybrid stacks—Burr for orchestration, Novyx for memory, and domain-specific tools for inventory and payments. Optimizing and launching a travel-planning app in minutes is now possible, but only if you accept that the LLM is a great conversationalist and a terrible accountant.
BCD's Model Context Protocol Play
The travel industry's agent frameworks have matured fast, but 2025's real question is which ones survive contact with production. Captain, the AI travel agent that surfaced on Show HN, makes a point worth repeating: LLMs are great, but they're not everything. Booking a trip requires deterministic logic, policy enforcement, and reliable state management layered on top of language models. Frameworks like Burr, the open-source full-stack agent framework, address this with explicit state machines that make agent behavior auditable — exactly what travel companies need when an itinerary goes wrong and someone has to explain why. Meanwhile, memory infrastructure like Novyx, with rollback, replay, and semantic search, solves the persistence problem that trips every demo-to-production pipeline.
The commercial layer is consolidating too. American Express publishing an Intent Intelligence Framework for trusted agentic commerce signals that payments rails are preparing for agents that book and pay autonomously. Pair that with real-time voice stacks like LiveKit Agents running on Oracle Cloud, and deployment timelines compress dramatically — launch a travel-planning application in minutes, not quarters. The winners in 2025 aren't the frameworks with the flashiest demos; they're the ones combining LLM flexibility with deterministic guardrails, durable memory, and commerce-grade trust.
Launching a Travel AI in Minutes
The 2025 landscape of AI travel agent frameworks has matured from experimental demos to production-ready platforms, and the difference between them comes down to orchestration, memory, and real-time interaction rather than raw model quality. Full-stack agent frameworks like Burr give developers the state management and control flow needed to build travel agents that can pause, recover, and follow business logic when an LLM inevitably goes off-script. Memory layers such as Novyx address the other half of the problem: a traveler's preferences, past itineraries, and mid-booking context need persistence with rollback and semantic search, because a flight-rebooking conversation that loses state is worse than no assistant at all. Meanwhile, Amex's intent intelligence framework signals that payment networks are formalizing how agents transact on users' behalf.
The travel-specific angle is what turns these primitives into something shippable in minutes. Platforms like getmtp's Captain show that the winning pattern is combining LLMs for natural conversation with deterministic tools for flights, hotels, and pricing, since hallucinated gate numbers or phantom availability destroy trust instantly. Add real-time voice infrastructure from frameworks like LiveKit Agents, and you get agents that can actually call airlines or handle inbound support at scale. The takeaway: model choice matters less than the framework's ability to constrain, remember, and execute reliably.
Top AI Travel Agent Frameworks Compared
| Framework | Strengths | Best For |
|---|---|---|
| Captain | Purpose-built travel agent with booking-aware workflows | Teams shipping travel-specific assistants fast |
| Burr (OS) | Full-stack, open-source agent orchestration with transparent state control | Developers who want full control over agent logic |
| Novyx + LLM stack | Memory API with rollback, replay, and semantic search for persistent context | Agents needing reliable long-term trip memory |
| LiveKit Agents | Real-time voice AI infrastructure, deployable on Oracle Cloud | Voice-first travel booking and support agents |