Loading ToolBench
MCP server for finding buying-intent conversations on Reddit: search, thread fetch, and two-phase lead scoring.
Three tools with complete schemas and descriptions. Naming is clear and verb-based (search_, score_, fetch_). Descriptions are detailed (100-200 chars) and explain WHAT, WHEN, and output. All parameters have types and descriptions. However, output schemas are not formally documented, responses are JSON text blobs without structured field definitions. Error handling is minimal; no recovery guidance or actionable error messages. The score_leads tool has a complex two-phase scoring mechanism that could benefit from clearer documentation of fallback behavior.
Fetch a Reddit post with its top comments, for reading the full conversation before drafting a reply.
Score Reddit posts as sales leads for a product using two-phase scoring: a lexical pre-score on every post, then Claude semantic scoring for posts above the escalation threshold (requires ANTHROPIC_API_KEY; falls back to the pre-score without it). Returns posts sorted by final score.
Search Reddit for recent posts matching a query, optionally restricted to specific subreddits. Returns post title, body, url, and engagement stats.
Output schemas not documented. Tools return JSON text blobs without formal field definitions. LLMs cannot plan downstream operations or extract specific fields reliably.
No error handling guidance. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next. E.g., Reddit API failures, missing ANTHROPIC_API_KEY, or invalid URLs return raw errors without recovery hints.
score_leads escalation_threshold parameter lacks clarity on fallback behavior. Description says 'falls back to the pre-score without it' (missing ANTHROPIC_API_KEY), but does not explain what happens if escalation_threshold is omitted or how the default 0.2 affects scoring.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
fetch_thread returns raw JSON with nested comment objects. No pagination or result limits documented. If a thread has 1000+ comments, the response could exhaust context windows.
search_reddit does not document what fields are returned in the post objects or how many results are typical. LLMs cannot reason about result structure or plan pagination.