Open Transfer Times

open-transfer-times-mcp

An MCP server that lets AI agents check whether a flight connection is realistic before recommending or booking it.

Flight-search tools return itineraries; they do not say whether a 50-minute layover in Frankfurt is enough. This server answers that with a deterministic, explainable assessment from the open-transfer-times engine, and it returns the cited airport facts behind every estimate so the agent can show its sources.

Runs locally over stdio. No API keys, no network calls.

Tools

All tools are read-only.

check_connection input

Give either both times or layover_minutes. If both are given they must agree.

Example result text:

BER–LHR–PVG: CAUTION — The terminals are not known, and a terminal change would leave little or no margin; ask for the terminals.Rating: REASONABLE — A practical connection on one ticket.Confidence: MEDIUM — The terminals are not known; a terminal change would make this connection tight or risky. With a terminal change (25 min) the margin would be 10 min (tight). The transfer-step durations are rule-of-thumb estimates, not sourced figures.Layover 90 min, transfer steps 55 min, margin 10 to 35 min (10 min if every possible step applies: tight).Steps counted: Move through airport 40 min (heuristic); Allow for another security check 15 min (heuristic). No other steps were counted.Possible step not counted: Change terminals 25 min (terminals not given). Give the margin as a range and ask for the terminals if it matters.Factors: Transfer through London Heathrow; Additional security screening may apply; Moderate transfer margin.Ticket: Protected itinerary.MCT: no published minimum connection time in this data. This is a structural estimate, not proof that an airline would sell this connection.Data: London Heathrow profile (partial, 3 cited facts).Comparing several options? Send them together to compare_connections, which ranks them.Airport facts are context, not extra steps. Apply a fact only where its scope matches this trip:- All connecting passengers go through security screening again at Heathrow.- ...

On separate tickets with bags not stated, where a re-check would matter:

BER–IST–PVG: INSUFFICIENT DATA — Separate tickets, and it is not known whether bags must be collected and checked in again; a re-check would make this connection tight or risky. With a bag re-check the margin would be -15 min (risky). The result cannot be relied on.Do not call this connection safe, fine or comfortable. Tell the user the estimate is unreliable, why, and what would settle it.Confidence: LOW, for the reasons above.Rough estimate only: layover 75 min, transfer steps 45 min, margin 30 min....Bags: not stated. If they must be collected and checked in again, the margin is -15 min (risky). Ask the user.

• The text leads with the decision, and compare_connections explains each option that is not recommended ("Why: …"). With insufficient_data the text leaves out the rating entirely: in the agent evaluation, agents repeated "comfortable" and "healthy margin" as reassurance even when the decision said the rating could not be relied on. The rating stays in structuredContent.
• mct is kept apart from the structural rating. below_published means the layover is under an airline's published minimum connection time at this airport (e.g. Lufthansa, 60 min at FRA); whether it applies to the trip is unknown, so it is never treated as a pass.
• decision is the field to act on. insufficient_data means key information is missing (the confidence reasons say what): ask the user rather than recommending the connection.
• steps always adds up to transfer_process_minutes. Facts are listed with their scope (for example "Air France connections at Paris airports") so the agent does not apply them to the wrong trip.
• Layovers are real elapsed time: at the 13 profiled airports a layover across a daylight-saving change is computed in the airport's time zone, and wall_clock_layover_minutes shows what the clocks say.

Instructions for agents

The server sends usage instructions when a client connects (MCP instructions): call a tool first with what the user gave, pass layover_minutes instead of inventing times, ask only when the airport or the duration is missing, use compare_connections for comparisons, explain only the counted steps, report the decision, ask about bags on separate tickets after giving the result, and respect fact scopes (scoped facts are shown as "(Only for: …)"). Each rule addresses a failure seen when testing agents; see eval.

The same data is returned as structuredContent matching the tool's output schema.

Setup

Requires Node.js 20+.

Claude Desktop — add to claude_desktop_config.json:

{  "mcpServers": {    "open-transfer-times": {      "command": "npx",      "args": ["-y", "open-transfer-times-mcp"]    }  }}

Claude Code:

claude mcp add open-transfer-times -- npx -y open-transfer-times-mcp

From a clone of this repository (before the package is published):

{  "mcpServers": {    "open-transfer-times": {      "command": "node",      "args": ["/absolute/path/to/open-transfer-times/packages/mcp/src/cli.js"]    }  }}

Run npm install at the repository root first.

Using it with a flight-search server

Pair it with any MCP server that searches flights. Ask the agent something like "Find flights from Berlin to Shanghai on 20 September and rank the one-stop options by how safe the connection is" — the agent searches with one server and passes the results to compare_connections.

Limits

• One-stop itineraries (two segments) only.
• A structural estimate from scheduled times and published airport information: not an official minimum connection time, and not a delay prediction.
• 13 airports have their own profile; all others use a standard baseline, which the result reports as coverage: "baseline".
• Clock changes are handled at the 13 profiled airports; elsewhere pass times with UTC offsets for exact layovers.

Development

npm install                          # at the repository rootnode packages/mcp/test/server.test.jsnpx @modelcontextprotocol/inspector node packages/mcp/src/cli.js   # interactive testing