Loading ToolBench
MCP server for Crisp customer support platform. Provides tools for interacting with Crisp's customer support platform including conversation discovery, messaging, state management, people/contacts lookup, and rich context assembly.
This server demonstrates solid naming conventions and reasonable descriptions for most tools. All 23 tools are explicitly registered with schemas and descriptions. However, there are consistent gaps in parameter descriptions, missing output schema documentation, and no error handling guidance. The tool set is well-organized for a customer support workflow (conversation discovery, messaging, people lookup, conversation management), but the implementation lacks production-grade error recovery patterns and advanced features like tool annotations or structured error messages.
Add segments (tags) to a conversation. Idempotent — existing segments are preserved.
Assign a conversation to an operator (by user_id), or unassign it (pass user_id: null).
Block a conversation (prevent the customer from sending new messages).
List conversations assigned to a specific operator (pass their user_id).
List unresolved conversations where the customer is currently waiting on an operator reply (unread_by_operator > 0). Ordered so the longest-waiting appear first. This is the single most useful 'what should I work on' tool.
List conversations tagged with a specific segment (e.g. `refund`, `bug`, `billing`).
Missing output schema documentation for all tools. Users/agents cannot determine what fields to expect from responses, forcing inference and risking parsing errors.
No error handling guidance. Tools like send_message, assign_conversation, and block_conversation are destructive/mutating but provide no recovery suggestions, classification (retryable vs user-fixable), or actionable error messages.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 64 | 2026-07-28+ | v2 |
| 2026-03-09 | D | 52 | - | v1 |
Look up a Crisp People profile by email address. Returns the profile if found, or `{}` if not.
Get detailed information about a specific conversation.
Get a conversation with messages formatted for analysis. Good for small-context summaries. For full customer history + cross-conversation context use `get_rich_context` instead.
Get message history for a conversation. File/animation/audio messages have their URLs surfaced in the output.
List all conversations for a Crisp People contact, across all of time. Returns a flat array of summarized conversation objects, sorted by most recent first.
Get the full Crisp People profile for a contact (by people_id). Returns comprehensive contact info: name, company, employment, geolocation, profiles, and more.
HEAVY CONTEXT: returns conversation + messages + linked Crisp People profile + that person's past conversations + custom data, all in one call. Replaces 4-5 round trips. Use this at the start of any non-trivial ticket reply.
Shortcut for all unresolved conversations across multiple pages. Returns a flat array (no pagination metadata). Use `conversations_awaiting_reply` if you specifically want ones where a customer is waiting on YOU.
List conversations with optional filters. Returns `{ data, pageNumber, hasMore, nextPage }` so callers know whether to fetch more. Filters are OR-combined with the standard search.
Mark all messages in a conversation as read by the operator.
Remove segments (tags) from a conversation. Idempotent.
Plain-text search across conversations.
Attach a file to a conversation by URL (no upload). Good when the file is already hosted (S3/R2/imgur/etc.). If you need to push raw bytes through Crisp's bucket, use the Crisp web UI or call the client's `uploadAndSendFile` method directly.
Send a text message or internal note to a conversation.
Change conversation state (pending / unresolved / resolved).
Unblock a conversation (allow the customer to send messages again).
Update conversation metadata (email, nickname, subject, segments).
Many parameter descriptions are minimal or missing detail. E.g., 'get_conversation' description is only 60 chars; 'set_conversation_state' lacks explanation of what states are valid; 'search_conversations' does not specify how plain-text search works or what fields are searched.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Agents cannot determine which tools are safe to retry, which have side effects, or which are read-only without LLM reasoning.
Pagination handling inconsistent. Some tools (list_conversations) include pagination metadata in the response structure description, while others (get_unresolved_conversations, conversations_awaiting_reply) accept max_pages but no documentation of response structure or whether results are incomplete when max_pages is hit.
Destructive operations lack confirmation mechanisms. block_conversation, unblock_conversation, and set_conversation_state (especially to 'resolved') are irreversible but have no dry-run or confirmation-request pattern.
Some parameters accept complex structures without clear constraints. E.g., 'segments' in add_segments and remove_segments are arrays of strings with no max length, pattern, or example guidance. LLMs may pass invalid segment names.