devresearch-mcp

# devresearch-mcp Let Claude Code read developer forums for you. When you ask *"is framework X worth using?"* — instead of Claude guessing from stale training data, it searches **Hacker News**, **Reddit**, and **Lobsters** in real time, reads what developers are actually saying, and answers from evidence. --- ## Install ```bash claude mcp add devresearch -- npx -y devresearch-mcp ``` Requires Node.js ≥ 20. **No API key. No config.** Restart Claude Code and you're done. --- ## Try it Ask Claude: > *"Use devresearch to check if Bun is production-ready"* Claude will: 1. Search Hacker News, Reddit, and Lobsters for Bun discussions (last 30 days) 2. Pull the top 6 threads with their comment trees 3. Score each post on 6 heuristics — velocity, expert engagement, dissent ratio, buzzword density, comment depth, longevity 4. Summarize: *"34 discussions found. Pro camp cites speed and drop-in compatibility; con camp flags native-module edge cases. Expert engagement is high (62%), buzzword density low — substantive discussion, limited hype."* --- ## What it can do | Ask Claude... | Tool used | |---|---| | *"What's changed in Tanstack Query v5?"* | `research` — aggregates threads + scores hype | | *"Search HN for WebAssembly discussions"* | `search` — keyword search across all platforms | | *"What's trending on HN today?"* | `trending` — front page / hot / hottest snapshots | | *"Show me HN thread 42"* | `get_post` — deep-dive with full comment tree | | *"What's jarredsumner's karma?"* | `get_user` — profile + karma lookup | You don't need to remember tool names — ask in plain English, Claude picks the right one. --- ## How hype detection works Each post is scored locally on 6 signals (no LLM calls): | Signal | What it measures | |---|---| | `velocity` | Score growth rate per hour since posting | | `buzzword_density` | Frequency of marketing words ("revolutionary", "10x", "game-changer"...) | | `expert_engagement` | Share of comments from high-karma authors | | `dissent` | Share of top-level comments expressing disagreement | | `depth` | Median comment-tree depth (deeper = more technical debate) | | `longevity` | How long the discussion stayed active | A post is classified as: - **strong_hype** — high buzzword + high velocity + low expert engagement - **mild_hype** — elevated buzzword and velocity with some expert presence - **substantive** — low buzzword + high expert engagement - **balanced** — signals don't clearly lean either way Claude then reads the raw data and writes the final narrative — it can cite specific comments, identify key voices, and flag misconceptions that the heuristics alone would miss. --- ## Cache Results are cached in a local SQLite database at `~/.devresearch-mcp/cache.db` with a 24-hour TTL. Re-running the same query within a day is instant and uses no network. --- ## Configuration (optional) Defaults work out of the box. To customize, create `~/.devresearch-mcp/config.toml`: ```toml [sources.reddit] subreddits = ["programming", "rust", "LocalLLaMA", "ExperiencedDevs"] [cache] ttl_hours = 24 [hype_scoring] buzzwords = ["game changer", "revolutionary", "next-gen", "paradigm shift"] ``` Override the config path with `DEVRESEARCH_CONFIG=/path/to/file.toml`. --- ## Limitations - Reddit uses the public JSON endpoints — subject to rate limits; occasional delays. - Cross-platform duplicate clustering uses URL normalization + title n-grams. Accurate for canonical links, weaker for text-only discussion threads. - `research` returns at most 6 top posts (`normal`) or 12 (`depth: "deep"`) to keep context usage predictable. - Heuristic signals are English-biased — buzzword/dissent detection relies on English keyword lists. --- ## License MIT