Loading ToolBench
Persistent memory + FTS5 search for Claude Code conversations. MCP server exposes ~/.claude/projects/ history as queryable memory for any MCP-compatible AI agent.
Claudex provides 9 tools with comprehensive descriptions and well-structured input schemas. All tools have descriptions (10-200 chars range), all parameters are typed, and most have helpful descriptions. The tool set demonstrates good composition, each tool has a single clear responsibility, and output fields are designed to support chaining. However, no explicit output schemas are documented in the source, and tool annotations (readOnlyHint/destructiveHint) are absent. Error handling guidance is implicit but not explicit in descriptions.
Get a condensed context snapshot for the current project. Call this at the START of every session to load project memory: recent sessions, favorited sessions, and project stats. This is your primary memory tool — use it before doing any work.
Get full content of a specific session. Returns all messages (user + assistant) in chronological order. Use this to load a session for detailed analysis or context.
Get an AI-generated summary of a session. Extracts the key topic, decisions made, code changes, and outcomes. Use this to quickly understand what was done in a session without reading all messages.
List all Claude Code projects on this machine. Shows which project is currently active.
List sessions for a project. Defaults to the current project. Pass projectId to target a different project. Use this when the user asks about previous conversations, or when you need to find a specific session.
Retrieve stored memories from the project. Search by namespace, type, or free-text. Returns high-priority memories and recent additions. Use this whenever you need context that was saved in previous sessions.
No explicit output schemas documented. Tool descriptions state what tools return conceptually (e.g., 'Returns all messages', 'Returns ranked results'), but no structured schema for response objects is visible in the source. LLMs cannot plan downstream field access without documented output structure.
Missing tool annotations (readOnlyHint, destructiveHint, idempotentHint). The source marks tools with 'Risk: READ_ONLY' or 'Risk: WRITE' in comments, but this is not translated into formal MCP tool annotations. LLMs cannot infer safety/idempotence without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | B | 76 | 2026-07-28+ | v2 |
Search conversations using full-text search. Defaults to the current project. Searches all messages, system prompts, and file contents across conversations. Returns ranked results with context snippets.
Store important information in persistent project memory. Use this to save decisions, architecture insights, codebase patterns, or current focus areas so they persist across sessions. Memory is tagged by namespace and type for easy recall.
Update metadata for a session: custom title, tags, notes, favorite status, or visibility. Use this to organize sessions, mark important ones, or add context-specific notes.
Error recovery guidance is absent from descriptions. None of the tool descriptions explain what to do if a call fails (e.g., 'If session not found, use list_sessions to verify the project ID'). Agents will lack context for retry strategies.
No confirmation step for write operations. store_memory and update_session_metadata modify persistent state but do not offer dry-run or confirmation. An agent could accidentally overwrite or delete important context.
search_conversations supports FTS5 syntax but does not validate or guide LLM on query syntax. Agents may pass invalid boolean operators (e.g., 'AND' without quotes). Description should say: 'For quoted phrases, enclose in double quotes. For boolean: AND/OR/NOT between terms. Wildcards: *'