MCP server for querying and replaying local Claude Code sessions.
Five tools with complete input schemas and descriptions, but descriptions are terse (avg 95 chars, below 194-char baseline). All tools are read-only and well-scoped. Naming follows verb_noun convention (list_, search_, diff_, replay_, regression_). Output schemas are partially documented in code but not formally declared in tool registration. No error handling guidance, no parameter constraints (enums), and no pagination support despite tools returning lists. Missing per-parameter descriptions in some cases (e.g., 'n' in list_recent_sessions lacks detail on max behavior).
Compare two Claude Code sessions by prompt, assistant text, and tool-call shape.
List recent local Claude Code sessions with first prompt, turn count, and tool-call summary.
Summarize reference session shapes for checking behavior after a CLAUDE.md change.
Replay one Claude Code session as structured user/assistant turns and tool calls.
Search local Claude Code sessions by prompt, assistant text, tool name, or tool input.
Output schemas not formally documented. Tools return complex objects (SessionSummary, diff results, regression metadata) but MCP tool registration does not declare return types. LLMs cannot plan downstream calls or extract fields reliably.
No pagination support. list_recent_sessions and search_sessions return unbounded lists. If a user has 500+ sessions, the response could exhaust context. No limit enforcement, no cursor/offset, no total count.
Parameter descriptions are minimal. 'n' in list_recent_sessions says 'Number of recent sessions to list (max 100)' but code enforces no max. 'query' in search_sessions lacks guidance on format (regex? substring? phrase?). Missing dependency hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 68 | 2026-07-28+ | v2 |
No error handling guidance. If a UUID is not found, the tool throws 'Session not found: {uuid}' but does not suggest recovery (e.g., 'Try list_recent_sessions to find valid UUIDs'). LLM has no next step.
Tool descriptions lack context on when to use each tool. 'Compare two Claude Code sessions' does not explain why an LLM would call diff_sessions vs replay_session. No guidance on prerequisites or typical workflows.