[mcp-analysis] MCP Structural Analysis - 2026-07-29

[github-actions[bot]]

Overview

Tested 15 representative GitHub MCP tools across 11 toolsets on 2026-07-29. Average usefulness: 3.27/5. Two tools were fully blocked by permission gaps (get_teams, list_repository_collaborators, both rating 1) and two more were quietly degraded by integrity-policy filtering (list_issues, search_issues, both rating 2) — pulling today's average below the 30-day mean of ~3.6. Best performers: list_pull_requests and the prompt-injected github_context block, both 5/5. This run also surfaced a pagination-parameter inconsistency: list_workflows uses snake_case per_page (unlike every other list tool's camelCase perPage) and silently ignored the requested limit of 1, returning ~30 full workflow objects instead.

Key Metrics

Metric	Value
Tools Analyzed Today	15
Total Tokens (Today)	16,055
Average Usefulness Rating	3.27/5
Best Rated Tools	list_pull_requests, github_context: 5/5
Worst Rated Tools	get_teams, list_repository_collaborators: 1/5
30-Day Data Points	173 (16 collection dates, 2026-06-30 → 2026-07-29)
30-Day Avg Daily Tokens	13,761
30-Day Rating Trend	Stable (3.61 → 3.59 first/second half)

Full Structural Analysis Report

Usefulness Ratings for Agentic Work

Tool	Toolset	Tokens	Rating	Assessment
list_pull_requests (fields-scoped)	pull_requests	175	⭐⭐⭐⭐⭐	fields param cuts an unscoped ~1320-tok response to 175 tok while keeping head/base/merge-actionable data
github_context (injected block)	workflow_context	75	⭐⭐⭐⭐⭐	Zero-MCP-overhead identity source; always correct fallback since get_teams/get_me are unusable
get_file_contents (README.md)	repos	12,400	⭐⭐⭐⭐	Clean resource envelope but unbounded — file keeps growing (contributor list now 285+ entries)
list_code_scanning_alerts	code_security	15	⭐⭐⭐⭐	Empty-state payload is minimal and clear; needs a 2nd call for severity:high
list_discussions	discussions	160	⭐⭐⭐⭐	Compact, cursor pagination works correctly, embeds category + author
get_label	labels	38	⭐⭐⭐⭐	Flat 4-field object, zero bloat
search_repositories	search	145	⭐⭐⭐⭐	All discovery signals in a flat shape
list_commits (fields-scoped)	repos	30	⭐⭐⭐⭐	fields=[sha,html_url] gives near-zero overhead when only identity is needed
list_workflows	actions	2,700	⭐⭐⭐	per_page=1 silently ignored — returned ~30 of 340 workflows
search_users	users	95	⭐⭐⭐	Only login/id/profile_url/avatar_url — needs follow-up call for profile depth
list_branches	repos	35	⭐⭐⭐	perPage=1 returns first-in-order branch, not necessarily default branch
list_issues	issues	60	⭐⭐	One matching issue was silently redacted by integrity policy — looks like zero open issues
search_issues	search	65	⭐⭐	Same silent-redaction problem as list_issues
get_teams	context	32	⭐	Permission not granted under read-only integration token (recurring 403-equivalent)
list_repository_collaborators	repos	30	⭐	Permission not granted; prior runs saw integrity-filtered [] instead

Schema Analysis

Tool	Type	Depth	Key Fields	Notes
github_context	markdown	1	actor, repository, workspace, workflow-run-id	Prompt-injected, not an MCP call
list_pull_requests	array	2	number, title, state, head, base	fields param dropped body/user/reviewers entirely
get_file_contents	object	3	content[].text(SHA), resource.uri, resource.text	No range/partial fetch support
list_workflows	object	2	total_count, workflows[].id/name/path/state	per_page (snake_case) ignored
list_discussions	object	2	discussions[].number/title/category.name, pageInfo, totalCount	Cursor pagination functional
list_issues / search_issues	array	1	(redacted)	Integrity-filtered items vanish with no in-band signal
get_teams / list_repository_collaborators	error	0	—	Permission-gated under this token

Response Size Analysis (30-day average)

Toolset	Avg Tokens	Tools Tested
repos	4,353	21
actions	1,941	16
code_security	1,787	16
pull_requests	1,651	16
issues	1,361	16
search	823	17
discussions	185	16
users	140	14
workflow_context	76	9
labels	46	16
context	37	16

30-Day Trend Summary

Metric	Value
Data Points	173
Collection Dates	16 (2026-06-30 → 2026-07-29)
Average Daily Tokens	13,761
Average Rating Trend	Stable (~3.6 across both halves of the window)

Recommendations

• High-value tools (rating 4–5): list_pull_requests/list_commits with explicit fields, github_context, get_label, list_code_scanning_alerts, list_discussions, search_repositories. Use fields/perPage aggressively — it is the single biggest lever for cutting response bloat without losing actionability.
• Tools needing improvement: list_issues/search_issues should surface redacted-item counts in-band (not just a side notice) so agents don't mistake "filtered" for "empty." list_workflows should honor per_page like its sibling tools, or the server should standardize on one pagination param name across all list tools.
• Context-efficient tools (low tokens, high rating): github_context (75 tok), get_label (38 tok), list_commits fields-scoped (30 tok), list_code_scanning_alerts empty-state (15 tok).
• Context-heavy tools: get_file_contents on README.md (12.4k tok, growing) — prefer get_repository_tree first and fetch narrower paths; list_workflows (2.7k tok due to ignored pagination).
• Permission gaps: get_teams and list_repository_collaborators remain structurally unusable for autonomous agents under the read-only integration token across every measured run since 2026-07-13 — treat both as permanently unavailable in this environment rather than retrying.

Visualizations

Response Size by Toolset

Usefulness Ratings

Daily Token Trend

Size vs Usefulness

Generated by 🔍 GitHub MCP Structural Analysis · age00 · 154.3 AIC · ⌖ 21.2 AIC · ⊞ 10K · ◷

• expires on Jul 30, 2026, 4:47 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 #49131.