[mcp-analysis] MCP Structural Analysis - 2026-04-14

[github-actions[bot]]

9 GitHub MCP tools analyzed across 6 days of historical data (Mar 30 – Apr 14). New finding today: 4 tools hit the installation rate limit (15,000 req/reset), the first rate-limit event in this series. Tools that did respond show consistent patterns: search_repositories and get_file_contents remain top-rated (5/5), list_discussions is the most token-efficient (155 tokens, 4/5), and list_code_scanning_alerts continues to be the most problematic (no pagination, historically dumps 138K+ chars).

Today's metrics: 9 tools tested · 4,630 total tokens · 3.44 avg rating · Rate-limit affected: 4 tools

Full Structural Analysis Report

Executive Summary

Metric	Value
Tools Analyzed Today	9
Total Tokens (Today)	4,630
Average Usefulness Rating	3.44/5
Best Rated Tool	get_file_contents / search_repositories: 5/5
Worst Rated Tool	get_me: 1/5 (persistent 403)
Rate-Limited Tools	4 (list_pull_requests, list_workflows, get_file_contents, list_code_scanning_alerts)
Historical Days	6 days (2026-03-30 → 2026-04-14)
Avg Daily Tokens (30d)	26,826

---

Usefulness Ratings for Agentic Work

Tool	Toolset	Rating	Assessment
get_file_contents	repos	⭐⭐⭐⭐⭐	Direct content + SHA for cache detection; immediately actionable
search_repositories	search	⭐⭐⭐⭐⭐	Minimal output mode default; perfect for discovery and routing
list_issues	issues	⭐⭐⭐⭐	Good structured data; body size highly variable (200–8,000 tokens)
list_pull_requests	pull_requests	⭐⭐⭐⭐	Comprehensive with base/head detail; body can balloon from quoted content
list_discussions	discussions	⭐⭐⭐⭐	Most context-efficient at 155 tokens; excellent cursor pagination
list_workflows	actions	⭐⭐⭐	per_page param ignored; 5 URL fields per entry adds unnecessary bloat
list_label	labels	⭐⭐⭐	No pagination; all 464 labels returned every call; grows with repo
list_code_scanning_alerts	code_security	⭐⭐	No pagination; alternates between empty (fine) and 138K+ char dumps
get_me	context	⭐	Persistent 403 error across all 6 test days; integration lacks user scope

---

Schema Analysis

Tool	Type	Depth	Key Fields	Notable
get_me	error	1	—	403 on every run; no user profile access
get_file_contents	object	2	sha, mimeType, text	SHA enables change detection without re-reading
list_issues	array	3	number, title, state, body, labels, user	Body is free-form; can contain diffs, tables, code
list_pull_requests	array	4	number, title, base, head, draft, merged	Nested repo objects in base/head add depth
list_workflows	object	3	id, name, path, state, url, html_url, badge_url	5 URL variants per entry; total_count useful
list_discussions	object	3	number, title, category, user, html_url	Compact; category embedded; pageInfo cursor
list_code_scanning_alerts	array	4	number, rule, severity, state, location	No per_page; full dump on bad days
list_label	object	2	id, name, color, description	Clean flat structure; no pagination offered
search_repositories	array	2	full_name, language, topics, stars, forks	minimal_output=true default is excellent

---

Response Size Analysis (Successful Responses Only)

Toolset	Avg Tokens (30d)	Notes
code_security	~17,000	Wildly bimodal: 5 tokens (empty) or 34,000+ (full dump)
repos	~3,200	Stable; README content drives most of the size
actions	~2,800	per_page ignored; always returns 30 entries
labels	~2,100	Grows linearly with label count (now 464)
issues	~1,500	Highly variable; depends on body content
pull_requests	~800	Body-dependent; Copilot PRs quote full issue body
search	~350	Very consistent; minimal_output=true works well
discussions	~155	Most stable and efficient tool in the suite
context	~15	Error only; no usable data

---

Tool-by-Tool Analysis (Today: 2026-04-14)

Tool	Toolset	Tokens	Schema	Rating	Notes
get_me	context	15	error	1/5	403 rate limit (also structural: integration lacks user scope)
list_issues	issues	1,800	array/3	4/5	Semantic refactoring report body (~1,800 tokens)
list_pull_requests	pull_requests	15	error	4/5	Rate limited; historical avg 3.8/5
list_workflows	actions	15	error	3/5	Rate limited; historical avg 3/5
get_file_contents	repos	15	error	5/5	Rate limited; historical avg 5/5
list_discussions	discussions	155	object/3	4/5	totalCount=4,928 (+150 from Apr 10)
list_code_scanning_alerts	code_security	15	error	2/5	Rate limited; historically dumps 138K+ with no pagination
list_label	labels	2,250	object/2	3/5	464 labels (was 462 on Apr 10); growing steadily
search_repositories	search	350	array/2	5/5	github-mcp-server: 28,854 ⭐ (+124), 3,969 forks (+32)

---

Notable Trend: Rate Limiting (New Finding)

First occurrence of installation rate limiting observed today (Apr 14). The GitHub API rate limit of 15,000 req/installation reset at 12:26 UTC. Four tools were affected when called after list_issues (which has a large body). This suggests that:

• Agents running many parallel tool calls may deplete the rate budget quickly
• The large list_code_scanning_alerts response (137K chars) is particularly costly per call
• Sequential tool testing strategy should account for shared rate budgets across all MCP consumers

---

30-Day Trend Summary

Metric	Value
Data Points	54 (9 tools × 6 days)
Days Covered	2026-03-30 to 2026-04-14
Average Daily Tokens	26,826
Discussions Growth Rate	~55 new discussions/day
GitHub MCP Server Stars	28,854 (+373 since Mar 30)
Labels Count	464 (growing ~4/day)
Code Scanning Alerts	Highly variable (0 or ~34,000 tokens)
get_me Status	403 on all 6 days — structural issue

---

Recommendations

High-value tools (4–5⭐) — use freely:

• search_repositories: Best context efficiency for discovery
• list_discussions: Most stable and compact; ideal for polling
• get_file_contents: Direct content with SHA caching support
• list_issues / list_pull_requests: Good data but budget for body variability

Tools needing care (3⭐):

• list_workflows: Request pagination support for per_page; eliminate 3 of 5 URL fields
• list_label: Add pagination support; 464+ entries is heavy for large repos

Avoid or use with caution (1–2⭐):

• list_code_scanning_alerts: Implement per_page pagination urgently; current behavior can consume entire context window in one call
• get_me: Structural 403 — integration needs user-level OAuth scope to be useful

Rate limit awareness:

• Multiple parallel MCP tool calls deplete shared installation rate budgets
• Stagger tool calls or use cached responses where SHA enables detection of unchanged content

---

Visualizations

Response Size by Toolset

Usefulness Ratings by Toolset

Daily Token Usage Trend

Token Size vs Usefulness

References:

• §24398162325

Generated by GitHub MCP Structural Analysis · ● 278.9K · ◷

• expires on Apr 15, 2026, 12:19 PM UTC

[github-actions[bot]]

This discussion has been marked as outdated by GitHub MCP Structural Analysis.

A newer discussion is available at Discussion #26420.