Loading ToolBench
MCP server bridging context between Claude Chat, Code, and Cowork
Acheron demonstrates solid tool definition quality with consistent naming patterns, comprehensive parameter schemas, and detailed descriptions. All 6 tools follow verb_noun convention (bridge_*). Schemas are well-structured with proper JSON Schema types, enums, and constraints. However, there are gaps in output schema documentation and error handling guidance that prevent a higher score. The tool set is well-composed with clear single responsibilities and good parameter chaining (e.g., bridge_search_context returns IDs usable by bridge_get_context). Descriptions are contextually rich (exceeding the 50-200 char baseline) and include when/why/how usage patterns, which is excellent for agent tool selection.
Permanently delete a saved context. Use when the user says "forget this", "delete that", "remove that note", "I don't need that anymore", or "that's outdated, remove it". Requires the context ID — use search or list first to find it.
Retrieve the full details of a previously saved context by its ID. Use this after finding a context via search or list, when the user wants to see the complete content of a specific saved memory. Typically used as a follow-up: "show me that decision", "give me the full details on that one".
Browse and filter all saved contexts. Use when the user asks "what have I saved?", "show me my decisions", "what do I have for this project?", "list my preferences", "what did we do recently?", "show everything tagged with...", or "what happened in Cowork?". Unlike search (keyword-based), this tool browses by category — filter by project, surface, type, tags, or date. Returns newest entries first.
Remember something for later. Use this when the user says things like "remember this", "save this", "note this", "keep this for later", "don't forget", or when an important decision, preference, or insight comes up that should persist across conversations. This saves context that will be available in ALL Claude surfaces (Chat, Code, Cowork) — even in future sessions. Use proactively when you recognize something worth remembering: a decision made, a user preference expressed, a lesson learned, a key file identified, or a workflow established.
Output schemas not documented. Tools describe input parameters comprehensively but do not define the structure of returned data (e.g., bridge_search_context returns results but schema of each result object, field types, and pagination metadata are not visible in the provided source).
Error handling lacks recovery guidance. No evidence in tool descriptions of what an LLM should do if a tool call fails (e.g., if bridge_delete_context fails because ID not found, should the agent retry, ask the user, or search first?).
bridge_status lacks clear output specification. Description states it shows 'a summary' but does not document which fields are in the response (e.g., total_count, by_surface, by_type, database_size, date_range structure?).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | B | 74 | 2026-07-28+ | v2 |
Search through everything that has been saved across conversations. Use when the user asks "what did I decide about...", "what do we know about...", "did I save anything about...", "find my notes on...", "what was that thing about...", or any question that might be answered by previously saved context. Also use proactively when the user asks a question that saved context might answer — check before saying "I don't have that information". Searches across all surfaces (Chat, Code, Cowork) and all projects.
Show a summary of all saved knowledge: how many contexts are stored, broken down by surface (Chat/Code/Cowork) and type (decisions, preferences, insights, etc.), database size, and date range. Use when the user asks "how much have I saved?", "give me an overview", "what's in my memory?", or "how big is my context database?".
Pagination parameters present but total count / next_cursor not documented. bridge_search_context and bridge_list_contexts both support limit/offset, but no evidence of whether responses include total_count or availability of next page.
No confirmation/dry-run pattern for destructive operations. bridge_delete_context is a permanent deletion but lacks a preview or confirmation step to prevent accidental data loss.