Bug Report: An assistant using the Strava MCP connector underestimated a user's training history

Daniel Griffin ·

Components: Strava MCP connector (list_activities), Claude (claude-sonnet-5)
Date observed: 2026-09-18
Severity: Moderate. The user got a wrong assessment of their fitness history, stated with confidence.
Status: Reported for discussion. This report doesn't claim a root cause.

User intent

The user asked for a marathon training plan based on their actual training
history. They are a former ultramarathon runner returning to training, and
wanted a plan that accounted for their long-term base rather than only their
recent weeks.

What happened

  1. The assistant retrieved activities since 2026-06-01. It found a longest
    recent run of about 5 miles and advised against a 10-week marathon build.
  2. The user asked: "But you see my historical base?"
  3. The assistant retrieved the 100 most recent activities, back to 2023-06-16.
    The response included has_next_page: true. The assistant reported:
    "Longest run ever logged: ~9.5 miles" and "he's never built
    marathon-distance mileage."
  4. The user pasted race results from UltraSignup, including a 200-mile finish
    in 2017.
  5. The assistant said those races predated the Strava account.
  6. A later date-filtered query (2015–2018) returned the 2017 data from the same
    account, including an activity titled "Recovery Run from Bigfoot 200."

Expected behavior

When a user asks about their training history, the assistant either retrieves
all of it or states plainly how much it retrieved.

Friction observed

In the connector

  • The tool description says nothing about how much of the history one call
    returns, or how to reach the rest.
  • A call with no arguments returns the 30 most recent activities.
  • Responses don't include a total activity count or the date of the oldest
    activity. The only sign that more data exists is has_next_page.
  • There's no single tool for retrieving the full history.
  • Strava's launch announcement promises access to users' "full training
    history." The help documentation says nothing about date ranges or
    pagination.

In the conversation

  • The assistant didn't tell the user which date range each query covered.
  • It stated conclusions about the user's whole history ("ever," "never") after
    retrieving only part of it.
  • It gave a reason for the missing data (that the account didn't exist yet)
    without checking it.
  • The pasted race results ran two fields together, so division place and
    finish time appeared as one string. The assistant read them without
    separating them and reported a 200-mile time of 46:20. The actual time was
    96:20:18.

Notes

  • The connector's date filters and pagination worked correctly when used.
  • Searching for similar public reports found none. The connector began rolling
    out the week of June 1, 2026.

This post was drafted by Claude (Opus 5) in a claude.ai session, at Daniel's direction and through several rounds of his revisions. The session it describes was run by Claude (Sonnet 5). Daniel directed the framing and editorial choices and takes responsibility for the published version.

Daniel Griffin