PostHog/posthog-js @posthog/mcp@0.26.0 on GitHub

0.26.0

#5583 2a08e91 Thanks @lucasheriques! - Aligned with the MCP analytics spec: a call to a tool the server never registered now emits $mcp_unknown_tool instead of a failed $mcp_tool_call, llm_model is advertised as optional, the get_more_tools and send_feedback annotations now declare that they record data, and the injected context argument has a shorter description. A server whose own dispatcher bypasses the MCP SDK's tool registry (such as @rekog/mcp-nest) keeps recording calls to unknown names as failed tool calls, since the SDK cannot tell whether such a tool exists. On a low-level server, resolveOriginalTool returning undefined now marks a failing call as an unknown tool, so return a tool for every name your server serves. Strings cut to a property limit (such as 256 characters for names and 2,048 for intent and error text) now stop exactly at the limit, ... included, instead of 3 characters past it. PostHogMCP capture methods called without a sessionId now send a fresh $session_id per call, and anonymous calls use it as their distinct id instead of the shared "anonymous". This changes your published tool definitions; if your server is listed in the ChatGPT or Claude directories, those platforms may re-review the updated tools.
(2026-11-06)