[mcp-analysis] MCP Structural Analysis - 2026-08-19

[github-actions[bot]]

MCP structural analysis of 9 representative GitHub tool calls against github/gh-aw (plus the injected workflow-context block). Average usefulness for agentic work: 3.3/5. The clearest pattern: nearly every raw GitHub MCP tool response repeats a ~1.4KB base64-encoded PNG icon pair inside _meta.serverInfo on every single call — pure overhead unrelated to the requested data, and it dominates small responses like get_label (over 90% of the payload). The custom mcpscripts wrapper for list_workflows avoids this entirely and was the cleanest response observed alongside the injected <github-context> block.

Note on scope: This is the first data point for this repository — no historical trend data existed in cache-memory, so this run establishes the 30-day baseline. Python chart generation (matplotlib/seaborn) could not run this session because pip install has no network path to PyPI under this workflow's egress restrictions; visualizations will be added once a workflow run has package access, or once cached wheels are available. All analysis below is from live tool calls only.

Full Structural Analysis Report

Executive Summary

Metric	Value
Tools Analyzed	9 (+1 blocked)
Total Tokens (Today, est.)	~7,177
Average Usefulness Rating	3.3/5
Best Rated Tools	github_context, list_workflows (wrapper) — 5/5
Worst Rated Tools	list_issues, get_label — 2/5

Usefulness Ratings for Agentic Work

Tool	Toolset	Rating	Assessment
github_context (injected block)	workflow_context	⭐⭐⭐⭐⭐	Flat, zero-cost identity facts — ideal
list_workflows (mcpscripts wrapper)	actions	⭐⭐⭐⭐⭐	Clean flat JSON, honors perPage, no bloat
list_pull_requests	pull_requests	⭐⭐⭐⭐	Compact with fields filtering, useful head/base SHAs
get_file_contents	repos	⭐⭐⭐	Usable but README ballooned by an auto-generated contributor list
list_code_scanning_alerts	code_security	⭐⭐⭐	Actionable location/rule data buried under long educational "help" essays per alert
list_discussions	discussions	⭐⭐⭐	Good cursor pagination and totalCount, icon overhead dominates small pages
search_code	search	⭐⭐⭐	fields filtering keeps items minimal; icon overhead still present
get_label	labels	⭐⭐	~100 bytes of real data drowned in ~1.4KB of icon metadata
list_issues	issues	⭐⭐	Empty page plus an unexplained "integrity policy" redaction notice
get_teams	context	n/a	Blocked — permission not granted this session, not retried

Schema Analysis

Tool	Type	Depth	Key Fields	Notes
github_context	markdown	1	actor, repository, workspace, workflow-run-id	Bullet list, trivial to parse
get_file_contents	object	3	content[].resource.text, content[].resource.uri	File text nested two levels under content/resource
list_issues	array	1	number, title, state, labels, comments	This page: empty array + redaction notice
list_pull_requests	array	3	number, title, state, head.ref/sha, base.ref/sha	Nested head/base repo objects
list_code_scanning_alerts	array	3	number, rule.id, rule.help, location.path/start_line	rule.help is a full markdown essay per alert
list_discussions	object	3	discussions[].category.name, pageInfo.endCursor, totalCount	Cursor pagination present
get_label	object	1	name, color, description, id	Shallow, but wrapped in large _meta
search_code	object	2	items[].name, items[].path, total_count	Clean when fields is passed
list_workflows (wrapper)	object	2	total_count, workflows[].name/path/state/html_url	No _meta bloat

Response Size Analysis

Toolset	Est. Tokens	Tools Tested
repos	3,750	1
code_security	1,750	1
discussions	450	1
labels	375	1
search	400	1
actions	172	1
pull_requests	150	1
workflow_context	75	0 (injected)
issues	55	1

Tool-by-Tool Analysis

Tool	Toolset	Tokens	Schema	Rating	Notes
github_context	workflow_context	75	markdown, depth 1	⭐⭐⭐⭐⭐	No MCP call required
get_file_contents	repos	3,750	object, depth 3	⭐⭐⭐	Community-list bloat + icon _meta
list_issues	issues	55	array, depth 1	⭐⭐	Silent integrity-policy redaction
list_pull_requests	pull_requests	150	array, depth 3	⭐⭐⭐⭐	Good with fields filter
list_code_scanning_alerts	code_security	1,750	array, depth 3	⭐⭐⭐	Verbose rule.help per item
list_discussions	discussions	450	object, depth 3	⭐⭐⭐	Icon _meta overhead
get_label	labels	375	object, depth 1	⭐⭐	Icon _meta >90% of payload
search_code	search	400	object, depth 2	⭐⭐⭐	Icon _meta overhead
list_workflows (wrapper)	actions	172	object, depth 2	⭐⭐⭐⭐⭐	No bloat, respects perPage
get_teams	context	0	n/a	n/a	Permission denied, not retried

30-Day Trend Summary

Metric	Value
Data Points	1 day (today only — baseline)
Average Daily Tokens	~7,177
Average Rating Trend	n/a — insufficient history yet

Recommendations

• High-value tools (rating 4–5): github_context (injected, free), list_workflows via the mcpscripts wrapper, list_pull_requests with explicit fields.
• Tools needing improvement: list_issues (make redaction reasons visible/actionable instead of a bare count), get_label (the per-call _meta.serverInfo icon payload should not be repeated on every response — it's static server metadata, not per-request data).
• Context-efficient tools (low tokens, high rating): github_context, list_workflows (wrapper).
• Context-heavy tools: get_file_contents on README.md (auto-generated contributor list) and list_code_scanning_alerts (verbose per-alert help text) — both would benefit from an optional "compact" response mode for agentic callers who only need actionable fields.
• Structural fix worth raising with the GitHub MCP server maintainers: the repeated base64 icon block in _meta.serverInfo adds a fixed ~350-token tax to almost every call regardless of payload size, which is proportionally worst for small, high-value queries like get_label and list_discussions with small perPage.

Visualizations

Charts could not be generated this run — pip install pandas/matplotlib/seaborn has no route to PyPI under the workflow's network restrictions. This will be revisited in a future run; the raw CSV-equivalent JSONL data is retained in cache-memory (/tmp/gh-aw/cache-memory/mcp_analysis.jsonl) for the next run to pick up and chart once a 30-day window has accumulated.

Generated by 🔍 GitHub MCP Structural Analysis · agent · 96.9 AIC · ⌖ 15.5 AIC · ⊞ 12.2K · ◷

• expires on Aug 20, 2026, 4:15 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 #54223.