1. Replica MCP overview
MCP (Model Context Protocol) is an open standard that lets AI assistants call outside tools. The Replica MCP gives your assistant a set of Replica tools, so it can look up the data behind Safety Hub, Travel Patterns, and Traffic and Speeds while it answers you.
Plain language: The assistant finds the place you named, picks the right Replica data, runs the query, and produces output.
Same access as Replica Studio: You sign in with your Replica Studio login and can query the released datasets behind the Studio applications, limited to the products and geographies your organization licenses.
When to use it. The MCP is best for quick answers, side-by-side comparisons, briefings and memos, customized visualizations, and questions that combine datasets (for example, speeding events per vehicle-mile). Use Studio when you want to work visually in easy-to-use, map-centric interfaces complete with preconfigured data visualizations, workflows, and saved work.
AI assistants can make mistakes. Your AI assistant chooses which Replica queries to run and writes the answers you see. An assistant can misread your request, choose the wrong data or filters, miscalculate, or state figures that do not exist in Replica's data. Replica is not responsible for the queries an assistant runs, how it interprets your prompt or the results, or the accuracy of what it produces. Always verify important information before you rely on it.
2. Connecting to the Replica MCP
What you need to get started
Replica MCP access: Your Replica account team turns this on for your organization. Contact your Customer Success Manager or [email protected]. Once it is on, you can query the same data tables and geographies your subscription already includes.
An AI assistant: Use one that supports custom MCP connectors, such as Claude, ChatGPT, Microsoft Copilot, or Cursor. Most require a paid plan. Setup steps for each are below.
Your Replica Studio login: You sign in through Replica, so there is no API key to manage, and your password is never shared with the assistant. If you are already signed in to Replica in your browser, sign-in may finish without showing the login page.
Approval to use that assistant: Query results are sent to your AI assistant, so check that your organization's AI policy allows it.
Claude (claude.ai and Claude Desktop)
Open Settings → Connectors and click Add custom connector.
Name it Replica and enter
https://mcp.replicahq.com/mcp. Leave the OAuth client ID and secret blank.Click Add, then Connect, and sign in on the Replica login page.
In a new chat, open the tools menu and make sure Replica is turned on.
Pro, Max, Team, and Enterprise plans support custom connectors; check your plan if you do not see the option. On Team and Enterprise plans, an owner may need to add the connector for the organization first; each member then clicks Connect and signs in. Connectors sync between Claude Desktop and claude.ai, so you only add it once.
Claude Code
Run this in your terminal:
claude mcp add --transport http replica https://mcp.replicahq.com/mcp
Then start Claude Code, run /mcp, select replica, and choose Authenticate.
ChatGPT
Open Settings → Apps & Connectors → Advanced settings and turn on Developer mode. On Business and Enterprise plans, a workspace admin may need to allow this.
Back in Apps & Connectors, click Create.
Name it Replica, enter
https://mcp.replicahq.com/mcpas the MCP server URL, choose OAuth for authentication, and create it.Sign in on the Replica login page when prompted.
In a new chat, choose the Replica connector from the + menu.
Menu names and steps can vary by ChatGPT plan and app version.
Microsoft Copilot
Copilot connects to MCP servers through Copilot Studio, which your Microsoft 365 administrator usually manages. Your administrator adds a Model Context Protocol tool to a Copilot agent with the server URL https://mcp.replicahq.com/mcp and OAuth sign-in. Email [email protected] for help with this setup. Menu names and steps can vary by Copilot Studio version.
Cursor and other MCP clients
Any client that supports remote MCP servers with OAuth sign-in can connect. In Cursor, add this to your MCP settings (~/.cursor/mcp.json), then sign in when Cursor prompts you:
{ "mcpServers": { "replica": { "url": "https://mcp.replicahq.com/mcp" } } }For other clients, add a remote (Streamable HTTP) server with the URL above and complete the Replica sign-in.
3. Using the Replica MCP
Getting started
Turn the connector on: Enable it for the chat, then ask your question. Saying "use Replica" the first time helps the assistant pick the right tools.
Name places with their state: Write "Austin, TX" rather than "Austin". Cities, counties, states, census tracts, ZIP codes, and metro areas all work.
Year, season, and day type: Name them if they matter. Otherwise the assistant uses the latest release and a typical weekday.
Follow-up questions: The assistant keeps the context, so "now just the top 5 corridors" or "same thing for weekends" works.
Pre-configured analyses
The Replica MCP includes ready-made analyses that run a full, multi-step workflow from a few inputs. In Claude, open the + menu in the chat box and choose Replica to pick one. In Claude Code, type / and look for the replica entries. Clients that do not support MCP prompts can get the same result by asking for the analysis in plain language.
Analysis | What you get | You provide |
Corridor safety brief | The riskiest road segments in an area, ranked, with the crashes on them | A place, how many segments, optionally a map area |
Compare areas on safety | A side-by-side safety scorecard for two or more similar places | Two or more places, counties, or states |
Pedestrian safety priority areas | Census tracts where pedestrian crashes, fast traffic, and car-free households overlap | A place, how many tracts |
Speed management candidates | Streets where traffic runs well over the posted limit, with traffic volumes for context | A place, how many streets |
Corridor deep dive | Volumes, trucks, speeds by hour, and driver behavior along one named street | A street and place, optionally a map area |
Travel profile | How, why, how far, and when people travel in an area on a typical day | A place |
Mode access by group | How mode choice differs by vehicle access, income, or race and ethnicity | A place and the comparison |
Freight profile | Which roads carry medium and heavy trucks, and where commercial trips end | A place |
Network volumes for a traveler group | The road segments that carry the most trips by a group (for example, low-income travelers or cyclists) | A place and the group |
Data limitations summary | A plain-language summary of what the data can and cannot support | Nothing |
Example questions
"Use Replica to break down 2023 pedestrian crashes in Austin, TX by severity and census tract."
"Which road segments in Fairfax County, VA have the most hard braking on weekdays?"
"Compare speeding per lane-mile across the counties in the Atlanta metro."
"What are the 20 highest-AADT road segments in Denver County, and what share is trucks?"
"Show the hourly traffic profile and 85th-percentile speeds on Colfax Avenue in Denver."
"What share of trips in San Francisco on a weekday are made by low-income households, by mode?"
"Show the trip distance distribution for residents of Travis County, TX."
Tips for reliable answers
Verify before you rely on it: AI answers can contain errors or figures that do not exist in Replica's data. Before you use a number in a decision, report, or presentation, ask the assistant which tools, filters, dataset, and release it used, and check the figure in Replica Studio or a download. Treat any number the assistant cannot trace to a Replica tool result as unverified.
Normalize before comparing: Compare areas of different sizes with rates, such as events per lane-mile or crashes per vehicle-mile, not raw counts.
Read the warnings: Replica's tools flag results that could mislead, such as missing data months or fatal-only crash sources. Ask the assistant to tell you about any warnings.
Drill in: Start broad (a county by road class), then ask about the specific segments or tracts that stand out.
4. Replica data available via the MCP
At a glance
Seasonal: a typical weekday and weekend day, two seasons a year (modeled)
Trips and travelers: Trips and people are available by mode, purpose, distance, time, land use, and traveler demographics.
Distributions: Trip distance, duration, and start time, traveler age, and income are available as bands and percentiles.
Road segment trip volumes: Each road segment has trip volumes, which you can filter or split by mode, purpose, or traveler group.
Annual: one release per calendar year, by road segment
Traffic volumes (modeled): Annual average daily traffic covers all vehicles and trucks, by direction, and hourly volumes are also available.
Speeds (observed): Average, median, 85th, and 95th percentile speeds are compared with the posted limit, and 15-minute speed profiles are also available.
Semiannual: two six-month releases per year (modeled)
Driver behavior events: Speeding, hard braking, rapid acceleration, phone handling, very high speeds (over 80 mph), and suspected collisions are available by area or road segment, with lane-miles for rates.
Driver behavior comparisons: Ready-made comparisons cover counties, cities, and states.
By calendar year, as reported (observed)
Crashes: Individual reported crashes and crash counts are available by severity, year, time, conditions, and modes involved (pedestrian, cyclist, motorcycle, and more), for any area. Sources vary by state.
Reference
Geographies: Look up any state, county, city, census tract, block group, ZIP code, or metro area across the US.
Release information: The assistant can list which seasons and years are available, so you can ask for a specific one.
Your subscription applies. What you can query depends on your organization's subscription. If a product or area is not licensed, the assistant tells you instead of returning data.
Data dictionary
Seasonal: trips and travelers (modeled)
Item | Details |
Releases | Two seasons per year, Q2 and Q4 (for example, 2026 Q2), from 2023 Q4 onward, nationwide. Each covers a typical weekday and, for most seasons, a typical weekend day. |
Population data | People counts, age, income, and traveler attributes need the season's population release, which can come after its trips. Ask the assistant which seasons have it. |
Counts | Trips, people, and trip takers (distinct people making trips) |
Values | Trip duration (minutes), trip distance (miles), vehicle-miles traveled, age, individual income, household income; as a sum, average, minimum, or maximum |
Trip attributes | Primary mode, trip purpose, previous trip purpose, origin and destination land use, origin and destination building use, vehicle fuel type, transit submode |
Traveler attributes | Race and ethnicity, sex, age, household and individual income, employment status, work from home, commute mode, vehicles available, resident or visitor, residence building type, language, education, housing tenure |
Break down by | Up to two of the attributes above or geographies (for example, mode by purpose, or trips by origin and destination tract) |
Distributions | Trip distance, duration, and start time; age; and individual and household income: bands plus average, median, and percentiles |
Road segment volumes | Trips on each road segment, filtered by mode, purpose, or traveler group, or split by mode, purpose, or distance |
Compared with Studio | Results use the nationwide release, so they can differ from a Travel Patterns Study built on a regional season. Name the release when quoting a number. |
Annual: traffic volumes (modeled)
Item | Details |
Years | 2019 and 2021 onward |
AADT | Annual average daily traffic for all vehicles, single-unit trucks, and combination trucks; each in total and, from 2023, by direction |
Hourly volumes | Vehicles per hour (0 to 23) on an average weekday, weekend, or single day of the week, in total and, from 2023, by direction |
Each year is released in versions. Compare years only within one version, since a version change marks a modeling or data vendor change.
Annual: speeds (observed)
Item | Details |
Years | 2021 onward |
Annual speeds | Free-flow speed; average speed; median, 85th, and 95th percentile speeds (2024 onward); and the posted speed limit and 85th percentile minus the limit (2025 onward) |
15-minute speed profiles | The same speeds for each 15-minute slot of a typical weekday, weekend, or single day |
Each year is released in versions. Compare years only within one version, since a version change marks a modeling or data vendor change.
Semiannual: driver behavior (modeled)
Item | Details |
Event types | Speeding (with a threshold you choose, such as 10 mph over), hard braking, rapid acceleration, phone handling, speeds over 80 mph, suspected collisions |
Measures | Events per day (a daily average over the months selected), plus the number of road segments, centerline miles, and lane-miles they occur on |
Break down or filter by | Geography, road segment, road class, posted speed, hour of day, month, weekday or weekend |
Comparisons | Pre-calculated comparisons of counties, cities, or states: event counts, events per lane-mile and per centerline mile, including vulnerable road user exposure |
Releases | Two per year: Q2 covers March to August and Q4 covers September to February. The newest release covers the whole US; older ones cover some regions. |
By calendar year: crashes (reported)
Item | Details |
Crash records | Date and hour, severity, injury and fatality counts, modes involved, roadway and environment conditions, contributing factors, location, and the reporting source |
Counts | Crashes, fatalities, serious injuries, injuries, pedestrian fatalities, cyclist fatalities |
Break down or filter by | Severity, year, hour, modes involved, weather, lighting, road surface, collision type, intersection, road classification, jurisdiction, and every tract, city, or county in an area |
Sources | The richest available source for the area: state crash data where Replica has it, otherwise FARS (fatal crashes only) |
Severity, time, intersection, and roadway attributes mean the same thing in every state. Other attributes, such as collision type and weather, use each state's own wording; ask the assistant to list the values in your area before filtering on them.
Reference: geographies and road segments
Every analysis is scoped to a geography, a bounding box on a map, or a list of road segments. The assistant converts place names to 2020 Census identifiers (GEOIDs) for you.
Level | Example |
State | 06 (California) |
County | 06037 (Los Angeles County) |
Place (city or town) | 0644000 (Los Angeles) |
Census tract | 06037101110 |
Block group | 060371011101 |
ZIP code (ZCTA) | 90012 |
Metro area (CBSA) | 31080 (Los Angeles metro) |
Road segments are identified by Replica network link IDs, which are shared by the seasonal and annual data. Each link carries one direction of travel between intersections, with its street name, road class, lane count, length, OpenStreetMap way ID, and opposite-direction segment.
5. Limits
AI plan usage: Every request and the results it returns count toward your AI assistant's usage, such as message limits, token allowances, or API credits, under your plan with that provider. Multi-step analyses and large results use more.
Chat results: Results in the chat are capped at 2,000 rows per request. For more, ask for a download: you get a link to a Parquet file with up to 250,000 rows of traffic and speed data, 150,000 crash records, or 100,000 rows of anything else. The link expires after 15 minutes, and downloads need download access on your account.
Area size: Requests for seasonal trip data and annual traffic and speed data are limited to roughly the size of the largest US county or metro. Split a state into counties.
15-minute speed profiles: Each request can include up to 20 road segments.
Time limit: A request that runs longer than about 4.5 minutes stops with an error. Narrow the area or the question and try again.
Crash comparisons across states: Each crash request uses one reporting source, so compare areas in different states in separate requests.
Not available yet: The MCP does not yet offer daily traffic, VMT, or origin-destination data; saving results to Studio; Scenarios; transit ridership by route; or custom areas drawn in Studio.
6. Privacy and security
Sign-in: You sign in with your Replica account through standard OAuth. The assistant never sees your password, and you can disconnect at any time from your assistant's connector settings.
Access enforcement: Replica, not the assistant, enforces access. Requests outside your licensed products or geographies are refused.
Read-only: The MCP cannot change data in Studio or anywhere else.
What your AI provider receives: Your AI provider receives the results of each request as part of the conversation, under your organization's agreement with that provider.
Request logging: Replica logs each request (your account, the tool and inputs used, and whether it succeeded) for support, security, and usage reporting.
Responsibility for AI output: Data returned by Replica's tools is provided under your organization's agreement with Replica. Your AI assistant, not Replica, interprets your prompts, chooses the queries, and writes the answers. Replica is not responsible for queries an assistant makes based on its interpretation of a prompt, for the accuracy of an assistant's output, or for data an assistant presents that does not exist in Replica's data. Your use of the assistant is governed by your agreement with its provider.
7. Troubleshooting
What you see | What to do |
"Not enabled for Replica MCP access" | Your account does not have MCP access yet, or it has expired. Email [email protected]. |
"Not licensed" for a product or area | Your subscription does not include that product or geography. The assistant can drop unlicensed areas and continue with the rest. |
The connector shows Disconnected | Click Connect in your assistant's connector settings and sign in again. |
The wrong place was used (for example, Austin, MN) | Add the state to the place name. |
"The requested area is too large" or a time-limit error | Use a smaller area, fewer years, or a narrower question. |
An empty result after filtering | Ask the assistant to list the valid values for that filter in your area, then try again. |
Access changed but the assistant does not see it | Wait 5 minutes and ask again. Replica refreshes access every 5 minutes. |
An answer with no Replica source, or the assistant did not use Replica | Turn the connector on, say "use Replica" in your question, and ask which Replica tool produced each number. Treat any number without one as unverified. |
"View-only access" when downloading | Your account can view this data but not download it. Contact your Customer Success Manager for download access. |
8. Getting help
Email [email protected] with questions, access requests, or feedback on answers. When reporting a problem, include the question you asked, the assistant you used, and roughly when you asked.
About Replica
Replica provides data about the built environment and how people move through it. We combine vast, disparate sources into a complete, up-to-date picture of mobility, land use, people, and economic activity, so planners and engineers can make decisions with current, consistent data instead of counts that are years old or incomplete.
Customers get insights via Replica Studio, our web platform. Its applications include, but are not limited to:
Travel Patterns Explorer: explore seasonal mobility data, including origins, destinations, modes, purposes, routes, and traveler demographics.
Traffic and Speeds Explorer: analyze annual traffic volumes and speeds along major roads across the country.
Daily Traffic and OD Explorers: nationwide, daily vehicle volumes on every road segment, vehicle miles traveled, and origin-destination trips for private auto, commercial, and on-demand travel, refreshed weekly.
Safety Hub: reported crashes and modeled driver behavior (speeding, hard braking, phone use) on every road.
Some technical users take advantage of Replica's Direct Database Access program via Google BigQuery. The Replica MCP provides a new way users can access Replica data.