[mcp-analysis] MCP Structural Analysis - 2026-08-17

[github-actions[bot]]

Ran a structural + size analysis across 12 GitHub MCP tool calls (one representative per toolset, plus one substitution and one supplementary probe). No historical data existed yet, so this is day 1 of a 30-day rolling trend.

Key findings: average usefulness 3.25/5 across tools tested. Best: github_context (workflow-injected block, 5/5) and list_workflows via the mcpscripts wrapper (5/5, clean pagination, no boilerplate). Worst: get_teams (blocked — permission not granted to this session) and search_code (hit a GitHub secondary rate limit, HTTP 429). The single biggest structural inefficiency found: nearly every built-in GitHub MCP response embeds a ~625-token-per-icon base64 PNG blob in _meta.serverInfo.icons (light+dark), which is pure branding overhead repeated on every call — for small lookups like get_label this overhead was ~98% of the response size.

Full Structural Analysis Report

Executive Summary

Metric	Value
Tools Analyzed	12
Total Tokens (Today)	10,450
Average Usefulness Rating	3.25/5
Best Rated Tool	github_context / list_workflows (wrapper): 5/5
Worst Rated Tool	get_teams / search_code: 1/5

Usefulness Ratings for Agentic Work

Tool	Toolset	Rating	Assessment
github_context	workflow_context	⭐⭐⭐⭐⭐	Excellent — zero-cost identity context
list_workflows (mcpscripts wrapper)	actions	⭐⭐⭐⭐⭐	Excellent — clean pagination, no bloat
list_pull_requests	pull_requests	⭐⭐⭐⭐	Good — minimal, fields-filtered
list_code_scanning_alerts	code_security	⭐⭐⭐⭐	Good — unambiguous empty-state signal
search_issues	search	⭐⭐⭐⭐	Good — substituted for rate-limited search_code
list_discussions	discussions	⭐⭐⭐⭐	Good — solid cursor pagination shape
get_file_contents	repos	⭐⭐⭐	Adequate — no range/excerpt support, full-file only
list_issues	issues	⭐⭐⭐	Adequate — silent integrity-filter redaction on one item
get_label	labels	⭐⭐⭐	Adequate — signal is ~2% of payload, rest is icon overhead
list_label	labels	⭐⭐	Limited — no perPage param, always dumps full label catalog (884 labels)
get_teams	context	⭐	Poor — blocked by ungranted permission, untestable
search_code	search	⭐	Poor — hit secondary rate limit (429), no structured retry signal

Schema Analysis

Tool	Type	Depth	Key Fields	Notes
github_context	markdown	1	actor, repository, workspace, workflow-run-id	Flat bullet list
list_workflows	object	2	total_count, per_page, page, workflows[]	Clean list+pagination envelope
list_pull_requests	array	1	number, title, state	Fields-filtered, flat
list_discussions	object	3	discussions[], pageInfo, totalCount	Cursor pagination (endCursor/hasNextPage)
get_file_contents	object	4	content[].resource.text/uri/mimeType	Resource-wrapper indirection adds depth
get_label / list_label	object/array	1	name, color, description, id	Flat but icon-overhead-heavy

Response Size Analysis

Toolset	Avg Tokens	Tools Tested
repos	6200	1
labels	1530	2
discussions	800	1
workflow_context	90	1
actions	175	1
issues	60	1
pull_requests	35	1
search	13	2
code_security	5	1
context	0	1 (blocked)

Tool-by-Tool Analysis

Tool	Toolset	Tokens	Schema	Rating	Notes
github_context	workflow_context	90	markdown	⭐⭐⭐⭐⭐	No MCP call needed
get_teams	context	0	error	⭐	Permission not granted this session
get_file_contents	repos	6200	object	⭐⭐⭐	No excerpt/range support; icon overhead
list_issues	issues	60	array	⭐⭐⭐	One item redacted by integrity policy
list_pull_requests	pull_requests	35	array	⭐⭐⭐⭐	Minimal, actionable
list_workflows (wrapper)	actions	175	object	⭐⭐⭐⭐⭐	Best pagination shape observed
list_code_scanning_alerts	code_security	5	array	⭐⭐⭐⭐	Clear empty state
list_discussions	discussions	800	object	⭐⭐⭐⭐	Icon tax dominates small requests
get_label	labels	760	object	⭐⭐⭐	98% overhead-to-signal
list_label	labels	2300	array	⭐⭐	No pagination control at all
search_issues	search	25	array	⭐⭐⭐⭐	Substituted for rate-limited search_code
search_code	search	0	error	⭐	HTTP 429 secondary rate limit

30-Day Trend Summary

Metric	Value
Data Points	12 (day 1 — no prior history)
Average Daily Tokens	10,450
Average Rating Trend	Baseline established today

Recommendations

• High-value tools (rating 4-5): github_context, list_workflows (mcpscripts wrapper), list_pull_requests, list_code_scanning_alerts, list_discussions, search_issues — favor these projections and the fields/perPage filters when available.
• Tools needing improvement: list_label (add pagination — it currently has no way to scope below the full 884-label catalog), get_teams (session-level permission gate blocks the whole context toolset), search_code (no backoff/retry guidance surfaced when rate-limited).
• Context-efficient tools (low tokens, high rating): github_context, list_pull_requests, list_code_scanning_alerts, search_issues.
• Context-heavy tools: get_file_contents (6200 tokens, unavoidable without range support) and list_label (2300 tokens, no scoping available).
• Systemic finding: the built-in GitHub MCP server embeds ~625-token-per-icon base64 branding data in _meta.serverInfo on nearly every response. This is fixed overhead independent of payload size and disproportionately hurts small single-object lookups (get_label, list_discussions at n=1). Custom wrappers (like the mcpscripts list_workflows) that drop this metadata show meaningfully better tokens-to-signal ratios.

Visualizations

Response Size by Toolset

Usefulness Ratings

Daily Token Trend

Size vs Usefulness

Generated by 🔍 GitHub MCP Structural Analysis · agent · 113.3 AIC · ⌖ 8.26 AIC · ⊞ 12.1K · ◷

• expires on Aug 18, 2026, 4:13 AM UTC-08:00

[github-actions[bot]]

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

A newer discussion is available at Discussion #53673.