Loading ToolBench
Dual MCP server suite: rider-search-proxy (fronts JetBrains Rider's MCP server, summarizes/token-caps search responses) and gamedev-log-analyzer (parses game-engine/build logs with dedup, severity classification, and token-efficient extraction)
Strong domain-specific tool set for game engine log analysis with well-structured schemas and clear descriptions. All 14 tools have descriptions (avg 156 chars) and input schemas with typed properties. Naming follows verb_noun pattern (log_*). However, several parameters lack descriptions, and output schemas are not documented. Error handling guidance is minimal. Tool composition is excellent, each tool has a single, clear responsibility and outputs chain well (e.g., log_detect → log_search → log_locate). Security is good: no credentials exposed as parameters. Descriptions are LLM-optimized and include practical guidance (e.g., 'pair with rider-mcp-enforcer' in log_locate). Baseline: avg tool description 156 chars (within 10 - 1024 range); 100% of tools have descriptions; 86% of parameters have descriptions (12/14 tools fully described, 2 have minor gaps).
Show current effective gamedev-log-analyzer settings + config-file path.
Find editor log files (Unreal Saved/Logs, Unity Editor.log) for the project, newest first.
Compare two logs (A=base/before, B=new/after) and emit ONLY the delta: new errors, errors that disappeared, and groups whose count changed. Unchanged groups are omitted — token-cheap regression triage across runs. Without pathA/pathB it auto-picks the two newest detected logs (A=older, B=newest). Filters: query, severityMin, category, file, groupBy, minDelta.
Show or set Bash log-grep enforcement. Pass mode to set ('block' = deny raw grep/tail/cat over .log/.jsonl/Logs + nudge to gamedev-log; 'warn' = allow + nudge; 'off' = allow). Omit (or 'status') to show.
Extract ONLY chosen scalar fields from structured trace-log lines into a compact table — the biggest token win on dense per-frame logs. Field forms: `Key`, `Key.x|.y|.z`, `Key.Y|.P|.R`, `ts`, `dts`, `d:Key`, `step:Key` (deltas vs previous row).
Output schemas not documented. Tools return structured data (e.g., log_search returns grouped messages, log_locate returns file:line locations) but no schema is visible in the source. LLMs cannot plan downstream tool calls or extract fields without knowing the response structure.
Some parameter descriptions are minimal or missing. 'log_config' has empty input schema (no properties documented). 'log_tail' description is vague ('escape hatch'). 'log_learnings_reset' and 'log_savings_reset' lack context on what 'reset' means.
No error handling guidance. Tools do not document what errors can occur (e.g., file not found, invalid query syntax, permission denied) or how to recover. LLMs cannot self-correct on failures.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Report what the analyzer is learning from your logs (local): parse coverage, top categories, and templated shapes of UNPARSED lines (candidates for a new parser/category). Sanitized.
Clear the local learnings ledger.
Jump list: distinct `file:line` locations of matched entries only (no message bodies), ranked by severity then frequency — the most compact handoff for opening the offending source. Pair with rider-mcp-enforcer: feed the (basename) locations to find_files_by_name_keyword → read_file at the line. `basename: true` strips paths to filenames for Rider's name search.
Report how many tokens you've saved vs dumping raw logs into context (local, cumulative).
Clear the local savings ledger.
Search/analyze an editor log: parse severity + category + file:line, dedup repeated spam into templated groups with counts, severity-sorted + token-capped. Filters: query, severityMin, category, file. groupBy 'template' (default), 'callsite' (roll up by file:line), or 'code' (roll up by diagnostic code, e.g. C4996/LNK2019 — collapses noisy builds to one line per code).
Configure gamedev-log-analyzer (projectPath, logPath, …). Writes ~/.gamedev-log-analyzer/config.json; run /reload-plugins after.
Overview of an editor log: counts per severity + top categories. No message bodies.
Last N raw lines of a log (escape hatch).
Pagination not documented. Tools like log_search and log_fields accept 'max' or 'maxGroups' but do not specify default values, min/max bounds, or whether results are truncated. LLMs cannot reason about result completeness.
STDIO transport only. Server is not remotely accessible and cannot be used by hosted MCP clients. Limits deployment to local/CLI-only scenarios.