MCP: experiment-results-get response too large due to step_sessions data

Problem

The experiment-results-get MCP tool (via mcp.posthog.com) returns responses that are 80K+ characters, making it unusable for LLM-based MCP clients. The response gets truncated before the statistical summaries (probability, credible_intervals, significant) appear.

The root cause is that the raw API response includes step_sessions arrays inside each metric's variant data. These arrays contain individual user-level session records (event_uuid, person_id, session_id, timestamp) — thousands of rows per funnel step, per variant, per metric.

For an experiment with 7 funnel metrics, the step_sessions data for just one metric's control baseline consumes ~75K characters. The actual stats that MCP consumers need (probability, credible intervals, significance) appear after all the session data and are always truncated.

Example

Metric 0 baseline data starts with useful info:

baseline:
  key: control
  number_of_samples: 8945
  step_counts[6]: 1970,155,101,66,41,20
  step_sessions[7]:     <-- 75K+ of session data starts here
    - [100]:
      - event_uuid: 019c7179-...
        person_id: f89116cb-...
        timestamp: "2026-02-18 07:58:30.522000"
      ... (thousands more rows)

The statistical summaries (probability, credible_intervals, significant) come after ALL metrics' session data, so they never make it through.

Suggested fix

Strip step_sessions from the experiment-results-get MCP response before returning it to the client. These session-level records are useful for PostHog's UI ("View recordings" links) but not for programmatic/MCP consumers.

Alternatively, add an optional parameter like include_sessions: false (defaulting to false).

The fields MCP consumers actually need are:

• step_counts (funnel counts per step)
• number_of_samples (exposure count per variant)
• probability (chance to win)
• credible_intervals (confidence bounds)
• significant (boolean)
• expected_loss (if available)

Note: the PostHog/mcp repo is archived, so filing here instead.

Start by locating the backend handler for the experiment-results-get MCP tool, using the tool name and mcp.posthog.com reference. Confirm where step_sessions enters the response and check the relevant experiment or MCP tests, if available; done means session records no longer make the response unusably large while the listed statistical fields remain available.