Exploring MCP Sessions

An MCP session is one agent run: the set of $mcp_tool_call events sharing a $session_id, ordered by timestamp.

Listing sessions, reading a session’s tool calls, and summarising its goal each have a typed tool — reach for those first. Drop to HogQL only for the three things the typed tools genuinely can’t do (see When to drop to SQL). The full $mcp_* property schema and query recipes live in the shared reference: models-mcp.md (opens in a new tab).

Tools

The three mcp-analytics-* tools are gated behind the mcp-analytics flag and run the same code as the sessions UI, so results match the screen. If they aren’t in your tool list, the project doesn’t have the flag — fall back to posthog:execute-sql, which is ungated.

The date-window trap — read this first

The two detail tools default to a 7-day lookback. A session you found in a list that reaches further back will come back empty unless you pass its session_start as date_from:

• posthog:mcp-analytics-sessions-tool-calls — date_from is an absolute ISO timestamp; pass the session_start you got from posthog:mcp-analytics-sessions-list.
• posthog:mcp-analytics-sessions-generate-intent — same date_from query param, same reason.

Empty tool calls for a session that visibly exists is almost always this, not a data problem. Carry session_start forward from the list row.

Workflow: list recent sessions

posthog:mcp-analytics-sessions-list
{ "date_from": "-7d", "order_by": "-session_start", "limit": 100 }

Each row: session_id, tool_calls, session_start, session_end, tools_used, mcp_client_name, distinct_id (+ resolved person_email / person_name), and intent (empty until generated). Response is { results, has_next } — page with limit / offset.

Three sharp edges:

• order_by takes column names, not response field names. Sort call volume as tool_call_count (not tool_calls). duration_seconds sorts fine even though it isn’t returned. An unrecognised key silently falls back to newest-first — so verify the order you got is the order you asked for. Valid: session_id, session_start, session_end, duration_seconds, tool_call_count, mcp_client_name, distinct_id; prefix - to descend.
• There is no error filter and no error count on a session row. “Which sessions had errors?” is a SQL question — see below.
• distinct_id_count is always 0. The field is in the response but the backend never populates it, so don’t read it as “one distinct id per session” — it says nothing. To count distinct ids in a session, use SQL.

search does a case-insensitive substring match across session_id, distinct_id, mcp_client_name, and tools_used.

Workflow: read one session’s tool calls

posthog:mcp-analytics-sessions-tool-calls
{ "id": "<session_id>", "date_from": "<session_start>", "limit": 500 }

Chronological tool_name, intent, timestamp, duration_ms, is_error, error_message — read top to bottom to reconstruct the run. limit defaults to 500 (also the max), which is the whole page for almost every session; has_next tells you if more remain.

Caveat: tool_name here is the raw $mcp_tool_name. Unlike the tool-quality and tool-detail tools, this endpoint does not resolve the inner tool of a single-exec wrapper call, so wrapper calls show the wrapper. When the inner tool is what matters (comparing against a tool-quality ranking, tracing a specific tool through a run), use the SQL recipe below instead. The same applies to tools_used on the session list.

Workflow: summarise the agent’s goal

posthog:mcp-analytics-sessions-generate-intent
{ "id": "<session_id>", "date_from": "<session_start>" }

Summarises the session’s recorded $mcp_intent values via an LLM and persists the result; later calls return the cached summary. Returns { session_id, intent }. A 503 means LLM summarisation isn’t configured — fall back to reading the raw $mcp_intent values from the tool-call list.

When to drop to SQL

Four cases, all via posthog:execute-sql, which — unlike the typed tools above — is not gated behind the mcp-analytics flag.

1. The project doesn’t have the mcp-analytics flag. The typed tools simply won’t be in your tool list. Everything below still works; this query is the plain session listing:

SELECT
    $session_id AS session_id,
    min(timestamp) AS session_start,
    max(timestamp) AS session_end,
    dateDiff('second', min(timestamp), max(timestamp)) AS duration_seconds,
    count() 

2. Errored sessions. The session list can’t filter or count errors — add HAVING errors > 0 to the query above and order by errors DESC.

3. Effective tool names within a session — the coalesce the typed tool-calls endpoint doesn’t apply:

SELECT
    timestamp,
    coalesce(nullIf(toString(properties.$mcp_exec_tool_call_name), ''), toString(properties.$mcp_tool_name)) AS tool,
    toBool(properties.$mcp_is_error) AS is_error,
    toString(properties.$mcp_error_message) AS error_message,
    round(toFloat(properties.$mcp_duration_ms)) AS duration_ms
FROM events
WHERE event = '$mcp_tool_call'
    AND $session_id 

4. Cross-session aggregation — “sessions per day”, “sessions that used tool X and then failed”, custom breakdowns. Recipes in models-mcp.md (opens in a new tab).

Note $session_id is a materialised events column — the same id as $mcp_session_id. Reference it bare, never as properties.$session_id: the properties. accessor renders null-wrapped in SELECT but as the raw column in HAVING/ORDER, so a HAVING search would mismatch the GROUP BY key.

Constructing UI links

• Sessions list: https://app.posthog.com/project/<project_id>/mcp-analytics/sessions

Tips

• A session with many calls but no errors that ends abruptly often means the agent gave up — check whether the last call returned a large or empty result
• $mcp_intent is only present when the client supplied it; absence is common, so generate-intent is the more reliable goal signal
• To go from a failing tool (see exploring-mcp-tool-quality) to the sessions that hit it, search the session list by tool name — remembering tools_used holds raw names, so search the registered name, not the inner one

Related skills

• exploring-mcp-tool-usage — the front door: routes a broad “how is my MCP doing?” question to the right tool
• exploring-mcp-tool-quality — error rates and latency across all tools
• exploring-mcp-intent-clusters — group goals across many sessions