Mapbox gives AI agents live maps and more than 35 navigation controls

Why it matters

AI agents that schedule travel, dispatch vehicles or coordinate field work need current spatial data and permissioned controls, creating a new infrastructure layer beneath models. Mapbox is using its developer distribution and live telemetry network to compete for that layer against Google, HERE and Esri.

Mapbox, the location platform Eric Gundersen (@ericg) co-founded in 2010, on September 17th announced a package of location services for AI applications and autonomous agents, giving software tools access to live map data and controls that can change routes, search for places and act on conditions in the physical world.

CEO Peter Sirota is extending a bet that Gundersen made when he founded Mapbox: maps should be programmable infrastructure that other developers can shape, rather than a finished consumer product. Gundersen arrived at that thesis through international development work, where incomplete maps made basic field operations harder. Sirota's version is built for software that needs to take the next step itself.

"AI can explain almost anything, but it doesn't deeply understand physical location," Sirota said in Mapbox's announcement. His pitch is that an agent planning a delivery, rescheduling a meeting or changing a driver's route needs structured information about roads, travel time, entrances and business hours. A fluent answer is of limited use when the coffee shop is closed or the driver is stuck behind an exit-ramp backup.

Turning a map into an agent tool

The package centers on what Mapbox calls its agentic mapping engine. Mapbox says autonomous agents process live GPS traces, sensor signals, satellite imagery, street-level video, user feedback and anonymized movement data from more than 45,000 applications. The resulting layers feed the map data behind Mapbox APIs and software development kits.

That engine supports the Mapbox Places API, now in public preview. Mapbox says the database spans more than 250 million points of interest and combines standard business details with building footprints, pedestrian entrances, operating hours, photos and activity patterns. Stable place identifiers are intended to keep developer integrations intact when a business changes its name or closes.

The scale claim puts Mapbox directly beside Google, which also says its Maps grounding service draws on data from more than 250 million places. Matching place counts do not establish matching coverage, freshness or quality. For developers choosing a location provider, the useful comparison will come from local completeness, update speed, pricing and the controls each platform permits an agent to use.

Mapbox also released Traffic 2.0 across its Maps SDK, Navigation SDK, Directions API, Matrix API and traffic data products. Mapbox says the engine produces accurate estimated arrival times for 98% of trips, forecasts traffic up to 2.5 hours ahead and can incorporate new speed signals within seconds. Mapbox's example is unusually concrete: congestion in an exit lane should affect drivers taking the exit without distorting the estimate for traffic continuing straight.

The 98% figure remains a company benchmark rather than an independently tested comparison. The result will matter differently for a five-minute urban trip and a long-haul logistics route. Traffic 2.0 is broadly available, while several other pieces of the package remain in preview, require access or were pre-announced.

Sirota packages the map for software that can act

Sirota became Mapbox CEO in March 2021 after serving as senior vice president of engineering. Before Mapbox, he founded and ran Amazon Elastic MapReduce and led engineering at Quantcast, experience centered on turning large data systems into services developers can operate at scale.

That background is visible in the less theatrical parts of the BUILD release. The new Mapbox CLI lets developers and coding agents configure and deploy Mapbox services from a terminal. Demo Access removes the requirement to provide a payment method before receiving a test token, reducing the chance that an experiment produces an unexpected bill. Mapbox also introduced a Figma MCP connector that can render maps from natural-language instructions inside Figma products.

An Agent Toolkit exposes more than 35 controls across Mapbox's mapping and navigation SDKs. Agents can add a stop, inspect how traffic affects a meeting, move the map view, add points or retrieve route summaries. Mapbox is also bringing those tools into Notion, where Notion Agents can standardize addresses, track deliveries and coordinate field operations through a hosted Mapbox MCP server.

The controls are the consequential part of the release. Place search gives an AI model more context for an answer. Route modification and application controls let an agent change what a user or vehicle does next. That expands the map provider's role from supplying background data to defining which physical-world actions software can execute and under what constraints.

A 16-year mapping thesis meets the agent cycle

Gundersen's route to Mapbox started far from consumer navigation. After studying international development at American University and working in microfinance in Peru, he co-founded Development Seed and built technology for governments and humanitarian organizations. Projects involving the World Bank, the United Nations and Doctors Without Borders repeatedly required geographic data that existing tools could not easily turn into customized maps.

A 2009 assignment in Kabul sharpened the problem. Gundersen recalled finding official maps with little useful city detail beyond two connecting roads. Mapbox grew from tools Development Seed had already built to process and publish geographic data, turning a project requirement into a developer platform.

Sirota is applying that same infrastructure thesis to agents. The difference is the customer inside the workflow. Earlier Mapbox tools helped a developer place a customized map in an application. The latest package is designed so an AI system can query the map, interpret the result and operate parts of the interface itself.

Mapbox has spent heavily to reach this point. In September 2023, Mapbox raised a $280 million Series E led by SoftBank to invest in AI location services and automotive products. The BUILD package broadens that strategy beyond the car, targeting application developers, coding agents, workplace software and design teams. Mapbox's current valuation, revenue and pricing for the newly introduced services are not part of the announcement.

The competitive field has also moved. Google made Maps grounding generally available in the Gemini API in 2025 and has since expanded Maps tooling for agent workflows. HERE Technologies and Esri signed a three-year location analytics partnership in July 2026, explicitly targeting AI systems and agentic workflows.

Mapbox is betting that developers will want a dedicated, programmable location layer even when their chosen AI model comes from a platform with its own map. The Places API, live traffic engine and agent controls give Mapbox a credible product package for that contest. Performance claims such as 98% ETA accuracy will need evidence beyond Mapbox's own release, particularly when customers are allowing agents to make decisions that affect vehicles, workers and deliveries.

Gundersen built Mapbox because people in the field could not act on maps that were blank or inflexible. Sirota's challenge is the modern version of the same problem: keeping an autonomous system from taking confident action on an incomplete picture of the world.