[mcp-analysis] MCP Structural Analysis - 2026-04-27

[github-actions[bot]]

Tested 9 representative GitHub MCP tools across 9 toolsets. Average usefulness rating: 3.0/5. Best: get_label (5/5 — minimal and flat). Worst: get_me (1/5 — 403 in this integration) and search_code/list_code_scanning_alerts (2/5 — severely bloated). Total tokens consumed across all tool calls: 9,416.

Full Structural Analysis Report

Executive Summary

Metric	Value
Tools Analyzed	9
Toolsets Covered	9
Total Tokens (Today)	9,416
Average Usefulness Rating	3.0 / 5
Best Rated Tool	get_label: 5/5
Worst Rated Tool	get_me: 1/5 (403 error in this integration)
Most Token-Heavy	get_file_contents (README): 2,750 tokens
Most Efficient	get_label: 33 tokens

---

Usefulness Ratings for Agentic Work

Tool	Toolset	Rating	Assessment
get_label	labels	⭐⭐⭐⭐⭐	Minimal flat object, immediately actionable
list_workflows	actions	⭐⭐⭐⭐	Clean structure; per_page parameter ignored (returns 30)
list_discussions	discussions	⭐⭐⭐⭐	Most compact listing tool; includes pageInfo + totalCount
list_issues	issues	⭐⭐⭐	Full HTML-encoded body always included — ~1000 tokens per issue
list_pull_requests	pull_requests	⭐⭐⭐	Body included; nested base.repo/head.repo adds redundant depth
get_file_contents	repos	⭐⭐⭐	Full file text, no truncation option; appropriate for targeted reads
list_code_scanning_alerts	code_security	⭐⭐	rule.help embeds full CWE documentation (~1,500 tokens per alert)
search_code	search	⭐⭐	Repository sub-object includes 40+ URL templates per result item
get_me	context	⭐	403 forbidden in this integration context; not functional

---

Schema Analysis

Tool	Type	Depth	Key Fields	Notes
get_label	object	1	color, description, id, name	Flattest schema tested
list_discussions	object	3	number, title, category, user, pageInfo, totalCount	Pagination cursor included
list_workflows	object	2	id, name, path, state, html_url, total_count	Wraps array in {total_count, workflows[]}
list_issues	array	3	number, title, state, labels, user, body	body is full issue text, HTML-encoded
list_pull_requests	array	4	number, title, state, head, base, labels, user, body	base/head each nest a full repo object
get_file_contents	object	2	content, sha, name, mimeType, uri	Returns raw text; no line/size limit
list_code_scanning_alerts	array	4	number, rule.id, state, most_recent_instance.location	rule.help is ~1,500 tokens of docs per item
search_code	object	5	total_count, items[].name, path, sha, repository	repository sub-object has 40+ hypermedia URL fields
get_me	error	1	error_message	403 in GitHub Actions integration token context

---

Response Size Analysis

Toolset	Tokens	Tool Tested
repos	2,750	get_file_contents (README.md, 11KB)
code_security	2,000	list_code_scanning_alerts (2 critical alerts)
actions	1,958	list_workflows (returned 30 of 268)
issues	1,000	list_issues (1 issue with full body)
search	875	search_code (1 result with full repo metadata)
pull_requests	625	list_pull_requests (1 PR with body)
discussions	150	list_discussions (1 item)
context	25	get_me (error response)
labels	33	get_label (single label)

---

Key Findings

Bloat patterns identified:

1. rule.help in code scanning: Each alert includes full CWE documentation as a multi-paragraph help text. For a page of 30 alerts this could consume ~45,000 tokens — the entire context budget of some models. Agents should filter this field out or use the html_url to fetch details on demand.

2. Issue/PR bodies: Both list_issues and list_pull_requests always return the full body text. For discovery/triage tasks, agents only need number, title, state, and labels. Bodies should only be fetched per-item via get methods.

3. search_code repository object: Every search result item embeds a full repository hypermedia object with 40+ URL template fields (archive_url, assignees_url, etc.). These add ~300 tokens per result and are never needed by agents — the html_url alone suffices.

4. per_page not honored on list_workflows: Passing per_page: 1 returned 30 items. Agents cannot rely on this for context control with this tool.

---

Recommendations

• High-value tools (rating 4–5): get_label, list_discussions, list_workflows — use freely
• Use targeted get methods, not list for issues and PRs to avoid body bloat
• Code scanning: always pass severity: critical filter; strip or ignore rule.help field in agent processing
• Search code: only use items[].path and items[].html_url; discard the repository sub-object
• Most context-efficient pattern: list_discussions (150 tokens/page) > get_label (33 tokens) for metadata lookups

---

Visualizations

Response Size by Tool

Usefulness Ratings

Tokens by Toolset

Size vs Usefulness

---

30-Day Trend Summary

Metric	Value
Data Points	9 (first run — no prior history)
Total Tokens (Today)	9,416
Avg Rating	3.0 / 5
History	Baseline established; trends available from next run

References: §24994722878

Generated by GitHub MCP Structural Analysis · ● 320.3K · ◷

• expires on Apr 28, 2026, 12:35 PM UTC

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-04-28T12:35:59.097Z.

Closed by Workflow