mvanhorn/last30days-skill
AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web - then synthesizes a grounded summary
How last30days-skill is put together
This repo is an AI-agent 'skill' distribution project: the core product is a Markdown skill specification (skills/last30days/SKILL.md) that instructs AI agents (Claude Code, Codex, Cursor, Gemini CLI, etc.) to run a multi-source 'last 30 days' research pipeline across Reddit, X, YouTube, TikTok, HN, Polymarket, GitHub and more, scoring results by social engagement and synthesizing a brief. A secondary runtime is a Go MCP server (module mcp/, built with mark3labs/mcp-go) served over stdio and packaged as a .mcpb for Claude Desktop; it registers a single research tool, runs preflight checks, and executes an embedded engine whose version-namespaced per-user cache lives under internal/engine. Python is present only for developer tooling (pytest, coverage gate, towncrier changelog); there are zero runtime Python dependencies. Data flows from agent/user invocation through the skill spec or MCP tool into parallel platform queries using user-supplied API keys/browser sessions, then into engagement-scored synthesis.
Languages
Frameworks
Infrastructure
Major components
Skill specification (skills/last30days)
SKILL.md is the source-of-truth runtime spec defining the /last30days command, setup wizard, and multi-source search/scoring behavior executed by AI agents.
MCP server entrypoint (mcp/cmd/last30days-pp-mcp)
Boots an mcp-go stdio server named 'last30days' and registers the research tool with a build-time-stamped Version.
MCP tools layer (mcp/internal/tools)
Registers the research tool and performs preflight checks (preflight.go) before running research requests (research.go).
Engine (mcp/internal/engine)
Embeds assets (embed.go), extracts them (extract.go), and runs the research pipeline with a Version-namespaced per-user cache (run.go).
Vendored libraries (mcp/internal/engine/vendored)
Holds vendored third-party code bundled into the engine binary, excluded from coverage.
Plugin/extension manifests
.claude-plugin, .codex-plugin, .grok-plugin, gemini-extension.json, and .agents/plugins define installation across 50+ agent-skill hosts.
Hooks (hooks/hooks.json)
Declares agent lifecycle hooks for hosts that support them.
Changelog & QA tooling
Towncrier fragments in changelog.d/ generate CHANGELOG.md; pyproject.toml enforces a branch-coverage floor of 84%.
A quieter week focused on two new source integrations plus routine security-tooling updates.
5 weeksWeek of 2026-08-17
Week of 2026-08-17latest
A quieter week focused on two new source integrations plus routine security-tooling updates.
Week of 2026-08-10
A big week for new data sources: X/Twitter search arrived via Grok, Amazon buyer reviews were added, and Reddit gained a backup data route.
Week of 2026-08-03
Mostly maintenance: a security patch for a known vulnerability, several bug fixes, and automated dependency updates.
Week of 2026-07-27
A security-critical fix plus a wave of accuracy improvements and translated documentation.
Week of 2026-07-20
The busiest week by far: 105 commits of community-contributed fixes, release-pipeline repairs, and automated dependency bumps.
Over the past five weeks this project — a tool that gathers and summarizes recent activity from many online sources — has been steadily expanding where it can pull information from (Telegram channels, X/Twitter via Grok, Amazon reviews, Perplexity) while fixing a long list of reliability bugs so results stay accurate and on-topic. A major security fix in late July closed a serious vulnerability, and the busiest week saw over a hundred commits of cleanup, dependency updates, and community-contributed fixes. The overall arc is one of broadening data sources while hardening the tool's correctness, security, and release process.
Week by week
2026-08-17A quieter week focused on two new source integrations plus routine security-tooling updates.latest5 changes
Feature
Telegram public channels as an optional source
Users can now opt in to having the tool gather recent posts from public Telegram channels.
Feature
Perplexity search moved to the Agent API
The Perplexity-powered web search integration was migrated to Perplexity's newer Agent API for better results.
Chore
Version bumped to 3.21.1
A routine release bump packaging up the week's changes.
Fix
Fixed broken star-history chart in docs
The project's README chart showing GitHub stars over time was displaying incorrectly and now renders again.
Chore
Security and CI tools updated
Automated bots updated several code-scanning and supply-chain-security tools to their latest versions.
2026-08-10A big week for new data sources: X/Twitter search arrived via Grok, Amazon buyer reviews were added, and Reddit gained a backup data route.6 changes
Feature
X/Twitter search via Grok CLI
The tool can now search X (Twitter) using the Grok command-line assistant, with Grok kept off by default until users enable it.
Feature
Amazon buyer-signal reviews
Amazon customer reviews became a new evidence source, pulled through the Bright Data service.
Fix
X results stay on-topic with retry judging
A new retry mechanism double-checks that X search results actually match what was asked before including them.
Fix
Reddit fallback when primary API fails
Reddit data collection now falls back to an alternate service (Arctic Shift) when the main listing endpoint is unavailable.
Fix
Out-of-date evidence no longer ranks highly
Results older than the requested time window are now excluded from top rankings instead of cluttering them.
Fix
Grok sessions and Amazon budgets handled more honestly
Expired Grok sessions are reported as degraded rather than healthy, and multi-source runs keep a fair time budget for Amazon searches.
2026-08-03Mostly maintenance: a security patch for a known vulnerability, several bug fixes, and automated dependency updates.4 changes
Fix
Security patch for text-handling library
Upgraded a core software component to fix a publicly known security vulnerability (CVE-2026-56852).
Fix
Full report saved even in compact mode
Fixed a bug where using compact output mode caused saved reports to be incomplete.
Fix
Cleaner GitHub topic searches
Removed stray search filters the planner sometimes injected into GitHub queries, which had been skewing results.
Chore
Dependency and security-tool updates
Automated bots refreshed secret-scanning, code-analysis, and other development tools across the project.
2026-07-27A security-critical fix plus a wave of accuracy improvements and translated documentation.6 changes
Fix
Critical security hole closed
Blocked a serious vulnerability where a malicious project could run harmful commands by planting secrets in a startup file.
Docs
READMEs translated into six languages
Added localized documentation including Simplified Chinese and four others, then corrected translation errors found afterward.
Fix
Four independently verified bugs fixed
A batch of confirmed bugs affecting result quality was fixed together in one pass.
Fix
Hacker News comments no longer silently dropped
Fixed an overly strict filter that was discarding nearly all Hacker News comments from results.
Fix
Reddit rate-limit errors reported correctly
When Reddit blocks requests due to limits or permissions, the tool now says so instead of falsely reporting 'no results'.
Fix
Windows setup wizard and YouTube timeouts fixed
The setup wizard now works properly on Windows, and comparison-mode YouTube searches stopped timing out.
2026-07-20The busiest week by far: 105 commits of community-contributed fixes, release-pipeline repairs, and automated dependency bumps.6 changes
Fix
Release pipeline repaired
Fixed broken automation files so version releases and changelog checks run reliably again.
Fix
YouTube transcript rescue made transparent
When a backup service rescues a failed YouTube transcript fetch, it's now logged openly rather than hidden.
Fix
Explicitly chosen sources are respected
Community fixes ensure the tool honors exactly which sources a user selects instead of overriding them.
Fix
Synthesis survives truncated input
Long summaries no longer break when their underlying data gets cut short mid-stream.
Fix
Chrome cookie handling across profiles
Fixed reading browser cookies when Chrome has multiple user profiles configured.
Chore
Security and build tools refreshed
Dependabot merged a series of updates keeping the project's scanning and build actions current.
Dependencies and code review
Dependency advisories
Security Watch
No known advisories across 14 scanned dependencies.
No known advisories in the scanned dependencies.
Code review
No concrete code-level issues confirmed.
Get this report every week for your repos.
GitZoid learns each repo, reports what changed, and flags what needs attention. One flat price for the whole team.
$19 a month, flat · First 10 outputs free · No card required