Strong tool naming with action verbs (rt_* prefix is consistent). Descriptions are detailed and domain-specific, averaging 180-250 chars, well above the 34-char baseline minimum. All 10 tools have descriptions. Input schemas present for all tools with type declarations. However, several parameters lack descriptions (e.g., rt_ratings 'filters' object has no field documentation), and output schemas are not explicitly documented in the visible code. Error handling guidance is minimal, tools do not indicate what to do on failure or how to recover. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications. rt_sign_in and rt_auth_status are well-designed for the 60-second Claude Desktop timeout constraint, but lack explicit error recovery guidance.
RTINGS' own articles: methodology, how-tos, lineups, and explainers.
Poll the status of an in-progress sign-in started by rt_sign_in. Long-polls under Claude Desktop's 60 s tool-call limit.
One test's curve for one product: the measurement methodology and the result.
One product's review: measurements, verdicts, pros/cons.
Rank and compare products by test results and usage scores.
RTINGS' own best-of lists by use, size, budget and brand, with reasoning.
What RTINGS measures in a category: test definitions, units, hierarchy, usages. With no `group`, returns the group/category tree with per-group counts. Pass a group's `original_id` (or name) as `group` to get that group's leaf tests with their units and precision — a top-level CATEGORY id works too and returns every test beneath it in one call. **`find="input lag"` searches every test and usage on the bench** by name, by group path ("panel" finds Panel Technology's Sub-Type) and by a word test's values, in one call — use it instead of walking the tree when you know roughly what the test is called; several terms at once as `find="face, weight, battery"` (each hit says which terms it matched). Feature flags and specs ("bagel", "slice capacity", "USB-C") are tests too, so `find` locates them faster than walking the tree. `original_id` is the stable key everywhere in this server — names repeat across groups (headphones has three "RMS Deviation From Target"); a repeated name must be passed as its id or qualified as "Group/Name". `insider_only: true` marks a test *gate-able*, not gated. Whether it is actually served depends on the category; `rt_ratings` reports what came back.
Output schemas not documented. Tools return data but LLMs cannot see field structure, forcing them to guess what fields are available for downstream chaining. rt_ratings, rt_product, rt_graph, rt_search, rt_article, rt_recommendations lack explicit response schemas.
Parameter descriptions incomplete. rt_ratings 'filters' is an object with no field documentation; rt_schema 'bench' lacks detail on valid values; rt_article 'url_path' format is unclear (relative? absolute? with/without leading slash?).
No tool annotations. All tools are marked READ_ONLY or WRITE in comments, but MCP tool annotations (readOnlyHint, destructiveHint, idempotentHint) are absent. LLMs cannot infer safety properties from names alone.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | <=2025-11-25 | v2 |
Cross-silo live search for products by name or model.
Start an in-conversation sign-in: open a browser window on the user's screen and wait for them to sign in to RTINGS. Call rt_auth_status(wait_s=45) to wait for the result; do not call rt_sign_in again while one is in progress.
List RTINGS' categories and, per category, whether its numbers are answerable now. Call this first. `data_completeness` is derived from what actually came back unblurred on this machine's last SIGNED-OUT fetch of that category: `full` (values and scores served), `gated` (withheld without a membership), `partial`, or `unknown` (nothing fetched yet, or only fetched while signed in, which cannot say what anonymous gets). When the session is a signed-in Insider, `data.this_session.access` is `full` and the column does not apply. `has_paywall` is true for all 28 and tells you nothing. `silos=["tv"]` returns just those rows.
Error handling guidance missing. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next. rt_sign_in mentions timeout but no recovery path; rt_auth_status lacks guidance on 'could_not_check' outcomes.
Pagination not explicit. rt_ratings accepts 'limit' and 'offset' but does not document max result count, whether a total_count is returned, or how to detect end-of-results. rt_search 'count' parameter lacks min/max bounds.