Platform · AI
Agents that understand the boat.
Every 32°N app exposes typed Model Context Protocol (MCP) tools. An AI agent discovers those tools by introspection, registers its permissions, and runs in the same sandbox as everything else. Ask whether to leave Wednesday or wait for Friday and the agent reads the same forecast the weather app uses, alongside your boat's polars and draft, before answering.
No bespoke integrations. No chatbot pretending to know things it cannot see. The AI has access to exactly the data the apps have.
How it works
One gateway. All apps. All context.
AI calls go through the 32°N AI gateway — a platform service running on the hub. The gateway adds boat context, enforces capability checks, logs every request, and controls cost. Apps do not wire up their own AI integrations. They expose MCP tools and the gateway makes those tools available to agents.
The agent reads the current GFS and ECMWF forecasts from the weather app's MCP tools, pulls the boat's polars and current draft from BoatHub, checks the departure waypoint against the chartplotter's route, and compares the predicted VMG (velocity made good towards the waypoint) on both departure windows. It returns a recommendation with the reasoning — which routing model favours which window, what the trade-off is, and where the forecast agreement breaks down.
The AI gateway
Context, control, and cost.
AI calls do not go directly to a frontier model. They go through the 32°N AI gateway first. The gateway does three things that matter: it injects boat context so the model always knows what boat it is talking about, it enforces capability checks so agents cannot access data they are not permitted to see, and it gives me cost visibility per request.
Boat context, always present
Every AI request that goes through the gateway carries the boat's current state: position, course, speed, wind, the active route, and the boat's identity. The model does not need to ask for these — they are in the system context on every call.
This is what makes the difference between a general-purpose chatbot and an assistant that can give a useful answer about your specific boat in its current situation.
MCP tools, not raw data access
Agents do not get direct database access. They call typed
MCP tools — functions like get_current_forecast(lat, lon)
or get_boat_polars(). Each tool is defined by the
app that owns that data. The gateway capability-checks each
tool call against the agent's registered permission set.
The audit log records every tool call: which agent called it, what arguments were passed, and what was returned. If an agent behaves unexpectedly, there is a complete trace.
Anthropic Claude as the primary model
The 32°N gateway routes to Claude by default. It is the best available model for multi-step reasoning tasks like passage planning and anomaly detection — where the answer requires combining several data sources and explaining the trade-offs. The gateway is model-agnostic; the default is a configuration choice, not a hard dependency.
Local inference for offline and latency-sensitive tasks
The hub runs Ollama on Apple Silicon. Smaller models — 7B to 13B parameters — run locally at real-time speeds without a cloud round-trip. Anchor watch decisions, instrument anomaly detection, and anything that needs sub-second response time goes through the local path. Cloud calls are for the reasoning-heavy tasks.
What the AI does
Practical assistance, not a chatbot.
The AI has access to real boat data. That changes what it can usefully do. These are the concrete capabilities — not the aspirational list, the things that work now or are being built.
Departure window analysis
Given a route and a set of forecast models, the agent compares departure windows, explains the trade-offs in plain terms, and flags where the forecasts disagree. It uses the boat's polars to estimate passage time under each scenario — not a generic assumed speed.
Swing and drag decisions
The anchor watch app detects drag by comparing GPS position to the set-anchor position over time. The AI layer adds context: wind shifts, tidal changes, surrounding boat traffic from AIS. It distinguishes normal swing from genuine drag and reduces false alarms.
Engine and systems monitoring
Engine temperature running 4°C above normal for the last two hours, with increasing coolant-level trend. The AI surfaces this as a pattern worth investigating — not just a threshold alarm. It has the maintenance log context: when the impeller was last changed, whether this engine runs warm under load.
Ask about the boat
"When did I last change the primary fuel filter?" — pulls from the maintenance log. "What is the tidal stream at Alborán at 06:00 tomorrow?" — reads the tide engine data. "Where is my nearest safe anchorage given current fuel?" — reads the chart and the tank levels. The AI knows the boat because it has the MCP tools.
Where this is going
Version by version.
The AI layer is being built incrementally. Read-only access first, then write access with explicit per-call permissions. Local model deployment follows once the permission model is solid.
Bundled assistant, read-only tools
The built-in assistant has read access to the weather app, the chartplotter, the boat data, and the maintenance log. It can answer questions about the boat's state and help with passage planning. All calls go through the AI gateway with full logging.
Write tools with explicit per-call permissions
Write tools: set an autopilot waypoint, start or stop an app, add a maintenance record. Each write tool call requires an explicit user confirmation displayed on the device — not a blanket permission grant at setup. The audit log records every write with the confirmation event.
Bring your own agent
Third-party agents — Claude in your own Anthropic account, a locally-hosted open model, an agent you write yourself — can connect to the 32°N MCP server and use the same tools the built-in assistant uses. Same sandbox, same audit log, same capability model. The permission system is the same regardless of whose agent is running.
See the platform in the demo.
The demo lets you explore the chartplotter and cockpit instruments. The platform page has the full five-layer architecture and links to every sub-page.