[mcp-analysis] MCP Structural Analysis - 2026-08-24

[github-actions[bot]]

Tested one representative tool per available GitHub MCP toolset against github/gh-aw (10 tools, ~17,960 tokens total). Average usefulness across all tools: 3.4/5. Best: github_context (5/5, zero-cost identity metadata). Worst: list_issues and search_issues (2/5 each) — both returned empty result sets because the only matching item was redacted by integrity policy, despite well-formed field-filtered queries. The biggest structural finding: list_label has no pagination parameter at all and returned all 902 repository labels (~4,900 tokens) in a single call.

Note

Data limitation this run: /tmp/gh-aw/cache-memory/ was mounted read-only in this session, so no historical data could be loaded or persisted — this report reflects a single-day snapshot only, not a 30-day trend. A missing_data signal was raised for this.

Full Structural Analysis Report

Executive Summary

Metric	Value
Tools Analyzed	10
Total Tokens (Today)	17,960
Average Usefulness Rating	3.4/5
Best Rated Tool	github_context: 5/5
Worst Rated Tool	list_issues / search_issues: 2/5

Usefulness Ratings for Agentic Work

Tool	Toolset	Rating	Assessment
github_context (injected)	workflow_context	⭐⭐⭐⭐⭐	Excellent — free identity metadata, no MCP call needed
get_label	labels	⭐⭐⭐⭐⭐	Excellent — tiny, precise single-label lookup
list_pull_requests	pull_requests	⭐⭐⭐⭐	Good — compact, field-filtered, embeds user info
list_workflows (mcpscripts)	actions	⭐⭐⭐⭐	Good — pagination metadata present, perPage respected
list_discussions	discussions	⭐⭐⭐⭐	Good — cursors + totalCount, category embedded
get_file_contents	repos	⭐⭐⭐	Adequate — usable but bloated by repo-specific content and repeated server icon metadata
list_code_scanning_alerts	code_security	⭐⭐⭐	Adequate — actionable location data, but duplicates a full rule-help tutorial per alert
list_issues	issues	⭐⭐	Limited — matching row redacted by integrity policy, empty result despite valid query
search_issues	search	⭐⭐	Limited — same integrity-redaction pattern, empty usable result
list_label (no pagination)	labels	⭐⭐	Limited — returns all 902 labels in one shot, no perPage/page param exists

Schema Analysis

Tool	Type	Depth	Key Fields	Notes
github_context	markdown	1	actor, repository, workspace, workflow-run-id	Flat bullet list, not a tool response
get_file_contents	object (resource)	3	content, resource.text, resource.uri	Wraps file text in an MCP resource envelope + serverInfo icons
list_issues	array	1	number, title, state, labels	Empty due to integrity filtering
list_pull_requests	array	2	number, title, state, html_url, user	user object embedded inline
list_workflows	object	2	total_count, workflows[]	Clean pagination envelope (total_count/per_page/page)
list_code_scanning_alerts	array	3	number, rule.*, location, most_recent_instance	rule.help duplicated verbatim per alert
list_discussions	object	2	discussions[], pageInfo, totalCount	GraphQL-style cursor pagination
list_label (all)	array	1	name, color, description, id	Flat but unbounded — no way to page
get_label	object	1	name, color, description, id	Flat, minimal
search_issues	array	1	number, title, state	Empty due to integrity filtering

Response Size Analysis

Toolset	Avg Tokens	Tools Tested
repos	8,100	1
labels	2,495	2 (all-labels vs single-label — huge spread)
code_security	3,100	1
discussions	380	1
pull_requests	320	1
workflow_context	350	1
issues	250	1
search	250	1
actions	220	1

Tool-by-Tool Analysis

Tool	Toolset	Tokens	Schema	Rating	Notes
github_context	workflow_context	350	markdown	5	Injected, zero MCP cost
get_file_contents	repos	8,100	object	3	README.md; size mostly from repo's own auto-generated contributor list, plus fixed serverInfo icon overhead
list_issues	issues	250	array	2	Empty — sole match redacted by integrity policy
list_pull_requests	pull_requests	320	array	4	Field-filtered, clean
list_workflows (mcpscripts)	actions	220	object	4	perPage correctly honored (unlike built-in snake_case-only tool)
list_code_scanning_alerts	code_security	3,100	array	3	2 open high-severity alerts; each carries a full duplicated CWE writeup
list_discussions	discussions	380	object	4	Good pagination + totalCount (9,956 total discussions in repo)
list_label (all)	labels	4,900	array	2	No pagination parameter — always returns all 902 labels
get_label	labels	90	object	5	Single-label lookup, minimal and precise
search_issues	search	250	array	2	Empty — sole match redacted by integrity policy

30-Day Trend Summary

Not available this run — /tmp/gh-aw/cache-memory/mcp_analysis.jsonl could not be read or written because the cache-memory mount was read-only in this session. Only today's snapshot (2026-08-24, 10 data points) is reflected above.

Recommendations

• High-value tools (rating 4–5): get_label, list_pull_requests, list_workflows (mcpscripts wrapper), list_discussions, and the injected github_context block — prefer these first when their data suffices.
• Tools needing improvement: list_label should support perPage/page — a repo with hundreds of labels currently forces a single oversized response. list_code_scanning_alerts and similar rule-based tools could omit the full rule.help tutorial after the first occurrence of a given rule ID to cut duplicate payload.
• Context-efficient tools (low tokens, high rating): get_label (90 tokens), list_workflows via mcpscripts (220 tokens), list_pull_requests (320 tokens).
• Context-heavy tools: get_file_contents on README.md (8,100 tokens, dominated by repo content, not protocol overhead) and unbounded list_label (4,900 tokens for 902 labels).
• Integrity redaction caveat: Both list_issues and search_issues returned zero usable rows this run because their only matching item was filtered by integrity policy. This isn't a schema/size problem, but agents relying on these tools should expect occasional silent gaps and not assume an empty array means "no matching issues exist."

Visualizations

Response Size by Toolset

Usefulness Ratings

Size vs Usefulness

References: §32724667110

Generated by 🔍 GitHub MCP Structural Analysis · claude · agent · 152.5 AIC · ⌖ 8.74 AIC · ⊞ 11.5K · ◷

• expires on Aug 25, 2026, 4:10 AM UTC-08:00