hn-mcp

HN threads have 500+ comments nested 10 levels deep. Your AI agent needs to read them without blowing its context window.

hn-mcp is an MCP server that gives AI agents full access to Hacker News — complete comment trees, search, and user profiles — with depth control so they can explore progressively instead of fetching everything at once.

Features

• Full comment trees — no depth limits, no truncation
• Depth control — fetch just top-level comments or the entire tree
• Smart pruning — reply_count at cut-off points so agents decide what to expand
• Search — full-text search across stories and comments with filters
• No API keys — uses the public Algolia HN API
• 100% test coverage — tested with VCR cassettes, no network calls needed

Typical agent workflow

1. get_thread(42123456, depth=1)       → story + 85 top-level comments with reply_counts
2. Agent picks Comment A (47 replies)
3. get_comment_tree(comment_a_id)      → full 47-reply subtree
4. Agent summarizes branch, picks next

Or for smaller threads, just get_thread(id, depth=-1) to get the entire tree at once.

Getting started

Standard config works in most tools:

{
  "mcpServers": {
    "hn": {
      "command": "uvx",
      "args": ["hn-mcp"]
    }
  }
}

claude mcp add hn -- uvx hn-mcp

{
  "mcpServers": {
    "hn": {
      "command": "uvx",
      "args": ["hn-mcp"]
    }
  }
}

{
  "mcpServers": {
    "hn": {
      "command": "uvx",
      "args": ["hn-mcp"]
    }
  }
}

claude mcp add hn -- uv run --directory /absolute/path/to/news-ycombinator-mcp hn-mcp

Prerequisites

• Python 3.12+
• uv (provides uvx)

Tools

Depth parameter

The depth parameter on get_thread and get_comment_tree controls how much of the tree you get:

depth=0   (no comments — story metadata only)
depth=1   Comment A (reply_count=3)       ← just the comment + count
depth=2   Comment A                        ← comment + direct replies
            ├── Reply A1 (reply_count=2)
            ├── Reply A2 (reply_count=0)
            └── Reply A3 (reply_count=1)
depth=-1  Full tree, no pruning

Development

git clone https://github.com/tomwojcik/news-ycombinator-mcp
cd news-ycombinator-mcp
make venv
make install

Run tests

make test

Tests use vcrpy cassettes — no network calls needed. Coverage is reported automatically.

Re-record cassettes

If the Algolia API response format changes:

# In tests/conftest.py, temporarily change record_mode to "new_episodes"
uv run pytest
# Then change it back to "none"

Project structure

src/hn_mcp/
├── app.py               # FastMCP instance + tree pruning helpers
├── client.py            # HNClient — async Algolia API client
├── server.py            # Entrypoint
├── types.py             # TypedDict definitions for all responses
└── tools/
    ├── get_thread.py
    ├── get_comment_tree.py
    ├── get_stories.py
    ├── search_stories.py
    ├── search_comments.py
    └── get_user.py

License

MIT — see LICENSE.

Contributing

Contributions welcome. Please open an issue first to discuss what you'd like to change.

When submitting a PR:

1. Add tests for new functionality
2. Record VCR cassettes for any new API calls
3. Ensure uv run pytest passes with 100% coverage