Skip to main content
Glama
hn-tech-signal-mcp

hn-tech-signal-mcp

by malkreide

πŸ‡¨πŸ‡­ Part of the Swiss Public Data MCP Portfolio

πŸ“‘ hn-tech-signal-mcp

Version License: MIT Python 3.11+ MCP No Auth Required CI

MCP server for global tech & AI signal intelligence β€” aggregates HackerNews, arXiv, Lobste.rs and GitHub into a structured briefing. No API key required.

πŸ‡©πŸ‡ͺ Deutsche Version

Demo

Demo: Claude using tech_signal_digest across arXiv, HackerNews and GitHub


Overview

hn-tech-signal-mcp turns any AI assistant into a proactive tech intelligence analyst. The server aggregates four signal layers β€” research frontier, developer discourse, curated signal, and open-source practice β€” into a single, structured briefing.

No authentication required. All four data sources are public APIs. Optional: set GITHUB_TOKEN for higher GitHub rate limits (5,000 req/h vs. 60 req/h unauthenticated).

Anchor demo query: "Give me a tech signal digest on AI today β€” what is happening in research, developer discourse and open source?"


Related MCP server: Knowledge Base MCP Server

Signal Architecture

FRONTIER    arXiv API      β†’ Latest AI/ML papers (cs.AI, cs.LG, cs.CL, cs.CV)
DISCOURSE   HackerNews     β†’ Six feeds + Algolia search + comment threads
            Lobste.rs      β†’ Curated, lower-noise tech signal
PRACTICE    GitHub Search  β†’ What engineers are actually building right now
            HN Show HN     β†’ What individuals are shipping this week

Think of the four layers as a radar: arXiv shows what's coming over the horizon, HN and Lobste.rs show what practitioners are discussing, and GitHub shows what teams are actually shipping.

Within the discourse layer there are two levels of depth. The feeds and the search tell you what is being discussed; hn_discussion tells you what is actually being argued β€” the counter-arguments and the "we tried this in production" replies that carry the real signal.


Features

  • πŸ”¬ Research frontier – Latest arXiv papers by category (cs.AI, cs.LG, cs.CL, and more)

  • πŸ” arXiv full-text search – Find papers by keyword, title, or author

  • πŸ—£οΈ HackerNews feeds – top, best, new, Ask HN, Show HN and YC job posts

  • πŸ”Ž HackerNews search – Full history via Algolia, with date range filter

  • πŸ’¬ HackerNews comment threads – Read the actual discussion under a story, nested, with a bounded fetch budget

  • πŸ”§ Lobste.rs hottest – Curated developer signal, filterable by tag

  • πŸ› οΈ GitHub trending AI repos – Search by topic, stars, sort by activity or popularity

  • πŸ“‹ Tech signal digest – One-call cross-source briefing in Markdown

  • ☁️ Dual transport – stdio for Claude Desktop, Streamable HTTP for cloud deployment

#

Tool

Source

Description

1

hn_top_stories

HackerNews

Six feeds: top/best/new/ask/show/job, with score filter

2

hn_search

HN Algolia

Full-text search across all HN history

3

hn_discussion

HackerNews

Nested comment thread under a story

4

arxiv_latest

arXiv

Latest papers by category (cs.AI etc.)

5

arxiv_search

arXiv

Search papers by keyword/title/author

6

lobsters_hot

Lobste.rs

Curated tech stories, filterable by tag

7

github_trending_ai

GitHub

Trending AI repos by topic and stars

8

tech_signal_digest

All sources

Aggregated Markdown briefing

HackerNews feeds

Feed

Content

Upstream size

top

Front page as ranked right now

500 items

best

Highest-voted recent stories

200 items

new

Newest submissions, unfiltered

500 items

ask

Ask HN β€” what practitioners are stuck on

~30 items

show

Show HN β€” what people are shipping

200 items

job

YC portfolio job posts (type: "job", no comments, score always 1)

~30 items

ask and job are short feeds upstream, so a large limit may return fewer stories than requested.


Prerequisites

  • Python 3.11+

  • uv or pip

  • No API key required

  • Optional: GITHUB_TOKEN for higher GitHub rate limits


Installation

# Recommended: uvx (no install step needed)
uvx hn-tech-signal-mcp

# Alternative: pip
pip install hn-tech-signal-mcp

Quickstart

# Start the server (stdio mode for Claude Desktop)
uvx hn-tech-signal-mcp

# With optional GitHub token for higher rate limits
GITHUB_TOKEN=ghp_yourtoken uvx hn-tech-signal-mcp

Try immediately in Claude Desktop:

"Give me a tech signal digest on AI today" "What are the latest cs.AI papers from the last 48 hours?" "What is HackerNews discussing about MCP this week?" "Show me trending GitHub repos for the topic 'ai-agents'"


Configuration

Environment Variables

Variable

Default

Description

GITHUB_TOKEN

–

Optional. GitHub personal access token. Without it: 60 req/h. With it: 5,000 req/h. The token is only sent to api.github.com, never to other upstreams.

MCP_TRANSPORT

stdio

Transport: stdio or streamable_http

MCP_HOST

127.0.0.1

Bind host for HTTP transport. Non-loopback values require MCP_BEARER_TOKEN.

MCP_PORT

8000

Port for HTTP transport

MCP_BEARER_TOKEN

–

Required when MCP_HOST is not loopback. Shared secret to gate the HTTP endpoint.

Claude Desktop Configuration

{
  "mcpServers": {
    "hn-tech-signal": {
      "command": "uvx",
      "args": ["hn-tech-signal-mcp"],
      "env": {
        "GITHUB_TOKEN": "ghp_yourtoken_optional"
      }
    }
  }
}

Config file locations:

  • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

  • Windows: %APPDATA%\Claude\claude_desktop_config.json

After restarting Claude Desktop, all 7 tools are available.

Cloud Deployment (Streamable HTTP)

For use via claude.ai in the browser (e.g. on managed workstations):

Render.com (recommended):

  1. Push/fork the repository to GitHub

  2. On render.com: New Web Service β†’ connect GitHub repo

  3. Optionally set GITHUB_TOKEN in the Render dashboard

  4. In claude.ai under Settings β†’ MCP Servers, add: https://your-app.onrender.com/mcp

# Local HTTP mode (binds 127.0.0.1 by default)
MCP_TRANSPORT=streamable_http MCP_PORT=8000 python -m hn_tech_signal_mcp.server

# Public bind (requires bearer token, intended behind a reverse proxy that terminates TLS)
MCP_TRANSPORT=streamable_http \
MCP_HOST=0.0.0.0 \
MCP_BEARER_TOKEN="$(openssl rand -hex 32)" \
python -m hn_tech_signal_mcp.server

Hardening: the server refuses to bind to non-loopback hosts unless MCP_BEARER_TOKEN is set. Run cloud deployments behind a TLS-terminating reverse proxy (Render, Fly, Caddy, …) and treat MCP_BEARER_TOKEN as the shared client secret your proxy enforces.


Architecture

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”    β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”    β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚  Claude / AI    │────▢│   HN Tech Signal MCP             │────▢│  HackerNews Firebase  β”‚
β”‚  (MCP Host)     │◀────│   (MCP Server)                   │────▢│  HN Algolia Search    β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜    β”‚                                   │────▢│  arXiv.org (Atom API) β”‚
                       β”‚  8 Tools                          │────▢│  Lobste.rs JSON API   β”‚
                       β”‚  Stdio | Streamable HTTP          │────▢│  GitHub Search API    β”‚
                       β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜    β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

Architecture decision

This server uses Architecture A (live API only, two paths per source). There is no bulk dump to fall back on.

Rationale (verified live on 2026-07-28 against the official HackerNews API):

  • All six feed endpoints ({top,best,new,ask,show,job}stories.json) answer HTTP 200 with 29–500 IDs. No auth, no rate-limit headers, Cache-Control: no-cache.

  • HackerNews publishes no bulk export, so caching is entirely this server's responsibility. TTLs live in CACHE_TTL.

  • The Firebase API has no search. Historical and full-text queries go through the Algolia index instead β€” that is the second path, used by hn_search.

  • item/<id>.json is one request per item. Feeds and comment threads therefore fan out, which is why both are bounded (HN_MAX_CONCURRENCY, max_comments).

Consequences:

  • Every upstream call retries with exponential backoff (2s / 4s / 8s) on network errors, 5xx and 429. Other 4xx fail fast.

  • One process-wide pooled httpx.AsyncClient, closed via the FastMCP lifespan.

  • Unknown item IDs return HTTP 200 with a null body rather than a 404 β€” hn_discussion translates that into an explicit "no item found" message.


Project Structure

hn-tech-signal-mcp/
β”œβ”€β”€ src/
β”‚   └── hn_tech_signal_mcp/
β”‚       β”œβ”€β”€ __init__.py
β”‚       β”œβ”€β”€ server.py          # All 8 tools
β”‚       └── outputs.py         # outputSchema models, one per tool
β”œβ”€β”€ tests/
β”‚   β”œβ”€β”€ __init__.py
β”‚   └── test_server.py         # 64 unit + 12 live tests
β”œβ”€β”€ pyproject.toml
β”œβ”€β”€ CHANGELOG.md
β”œβ”€β”€ CONTRIBUTING.md
β”œβ”€β”€ LICENSE
β”œβ”€β”€ README.md                  # This file (English)
└── README.de.md               # German version

MCP Protocol Version

This server speaks two protocol eras over the same endpoint. The client's first request on a connection decides which one applies; a later claim from the other era is refused.

Era

Revision

Who reaches it

initialize handshake

2024-11-05 … 2025-11-25

What today's clients speak. The server answers with the revision asked for, or with the 2025-11-25 ceiling when the request asks for something newer.

Per-request envelope

2026-07-28

A request carrying the 2026-07-28 _meta envelope opens a modern connection.

Both revisions are pinned in tests/test_protocol_version.py and asserted against the installed SDK, so a Dependabot bump of mcp cannot move either one silently.

Spoken, not just named: tests/test_modern_era.py actually connects β€” in-process in every era (2026-07-28, auto, legacy) and over HTTP with single POSTs, no initialize and no Mcp-Session-Id, against the ASGI app that MCP_TRANSPORT=streamable_http starts. It checks server/discover, tools/list with its freshness hint, tools/call with the Mcp-Name header, the serverInfo stamp in the _meta of every response, and that no feature deprecated by SEP-2577 (sampling, roots, logging) is involved. Two negative controls sit alongside: a wrong Mcp-Method header is rejected with -32020, and the same request under 2025-11-25 still requires a session.

Known, SDK-side: server/discover advertises resources and prompts (with listChanged / subscribe) although this server registers neither β€” MCPServer always installs those handlers. Switching that off would mean touching private attributes; instead the tests assert that both lists answer empty and without error.

Note that the SDK's LATEST_PROTOCOL_VERSION is an alias for the modern era, not for the handshake era β€” pinning against it alone would leave the era that current clients actually negotiate free to drift.

Update policy. When the gate fails, do not edit the constant blindly: read the spec changelog between the two revisions, verify the server still behaves, then move the constant, this section, README.de.md and CHANGELOG.md together.


Tool Output

Every tool declares its own outputSchema (closed: additionalProperties: false) from the models in src/hn_tech_signal_mcp/outputs.py, and answers in the three forms the spec separates:

Case

content[0].text

structuredContent

isError

Success

The JSON object, pretty-printed

The same object

false

Source failed

[Source] Error: …

β€”

true

Unknown or wrong ID (hn_discussion)

A sentence saying what to pass instead

β€”

true

Digest with some sources down

The JSON object, with degraded_sources and a per-source error

The same object

false

Before, the SDK derived a schema of {"result": string} from the -> str annotation, so structured clients got the JSON as one string field, and every failure arrived as isError: false.

The text stays byte-for-byte what it was β€” clients that only read text see no change. A response that does not fit its schema is reported as a server defect naming the offending field paths (never the values) instead of reaching the client unvalidated. The recorded fixtures and the daily live run both go through this path, in both protocol eras.


Testing

# Unit tests (no network required)
PYTHONPATH=src pytest tests/ -m "not live"

# Live integration tests (requires network)
PYTHONPATH=src pytest tests/ -m "live"

Example Use Cases

KI-Fachgruppe / AI Working Group

"Give me a tech signal digest on AI today"
β†’ tech_signal_digest(focus="AI")

"What are the top 5 arXiv papers on LLM agents this week?"
β†’ arxiv_search(query="LLM agents", category_filter="cs.AI", limit=5)

"What is HackerNews discussing about model context protocol?"
β†’ hn_search(query="model context protocol", days_back=30)

Research Monitoring

"Show me the latest NLP papers from arXiv"
β†’ arxiv_latest(category="cs.CL", limit=10)

"Search arXiv for papers on retrieval-augmented generation"
β†’ arxiv_search(query="retrieval augmented generation RAG", limit=10)

Open Source Intelligence

"What AI agent frameworks are trending on GitHub?"
β†’ github_trending_ai(topic="ai-agents", sort="updated", limit=10)

"Show me the most starred MCP-related repos"
β†’ github_trending_ai(topic="mcp", sort="stars", min_stars=50)

[β†’ More use cases by audience β†’](EXAMPLES.md)

arXiv Category Reference

Category

Full Name

Key Topics

cs.AI

Artificial Intelligence

Agents, planning, knowledge representation

cs.LG

Machine Learning

Training, optimisation, generalisation

cs.CL

Computation & Language

NLP, LLMs, translation, summarisation

cs.CV

Computer Vision

Image recognition, generation, multimodal

cs.RO

Robotics

Embodied AI, navigation

stat.ML

Statistics ML

Probabilistic methods, Bayesian ML


Rate Limits

Source

Auth Required

Limit

HackerNews Firebase

No

Very generous (Firebase)

HN Algolia Search

No

~10,000 req/hour

arXiv

No

~3 req/second (be respectful)

Lobste.rs

No

Reasonable use

GitHub Search

No

60 req/hour

GitHub Search

GITHUB_TOKEN

5,000 req/hour


Known Limitations

  • GitHub rate limit: 60 req/h without token. Set GITHUB_TOKEN for production use.

  • arXiv: Papers may take up to 24h to appear after submission. Weekends/holidays have delayed batches.

  • HackerNews: Top/best story lists update every few minutes. Very new stories may have low scores.

  • HackerNews ask / job feeds: Only ~30 items exist upstream, so a large limit returns fewer stories than requested. Job posts carry type: "job", no comment count, and a score of 1.

  • hn_discussion is always a sample, never the full thread: one request per comment upstream means popular stories (900+ comments) cannot be fetched whole. The budget is split across nesting levels and spread round-robin across sibling threads, so you get a representative cross-section rather than one exhaustively-read sub-thread. Check the truncated flag.

  • hn_discussion comment text is plain text, not HTML: HN's markup is stripped for readability. Do not re-render the output as HTML β€” the conversion is not a sanitiser.

  • Lobste.rs: Smaller community than HN; tech-focused but may not cover all AI topics.

  • tech_signal_digest: Makes ~4 concurrent requests; if one source is slow it may delay the full response.


Synergies with Other MCP Servers

hn-tech-signal-mcp combines well with:

Combination

Use Case

+ news-monitor-mcp

Global research + Swiss institutional media coverage

+ fedlex-mcp

Tech discourse + Swiss regulatory context

+ global-education-mcp

AI research trends + education policy data

+ swiss-statistics-mcp

Tech landscape + Swiss economic/structural data


Safety & Limits

  • Read-only: All tools perform HTTP GET requests only β€” no posts, comments, votes, or writes are issued upstream.

  • No personal data: The server queries public tech aggregators. No PII is collected; author handles in public posts/papers are returned as-is from upstream and never enriched or cross-referenced.

  • Rate limits: See the Rate Limits table. arXiv's ≀3 req/sec guidance is respected by default; GitHub search is capped at 60 req/h without a GITHUB_TOKEN. A request timeout is enforced per call.

  • No bulk harvesting: This server is built for interactive, conversational use β€” not for scraping or mirroring. Do not use it to bypass upstream pagination or ToS limits.

  • Terms of service: Data is subject to the ToS of each source β€” HackerNews, arXiv API, Lobste.rs, GitHub.

  • No guarantees: Community project, not affiliated with HackerNews / Y Combinator, arXiv / Cornell, Lobste.rs, or GitHub. Availability depends on upstream APIs.


Changelog

See CHANGELOG.md


Contributing

See CONTRIBUTING.md (Deutsch).


Security

See SECURITY.md (Deutsch) for the security posture and how to report a vulnerability.


License

MIT License β€” see LICENSE


Author

Hayal Oezkan Β· malkreide


Installation

Run via uv's uvx β€” no clone or manual install needed. Add to your MCP client config (mcpServers for Claude Desktop, Cursor and Windsurf; use a top-level servers key for VS Code in .vscode/mcp.json):

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

Available Tools

8 tools
arxiv_latestA
Read-onlyIdempotent

Fetch the most recently submitted papers from arXiv AI/ML categories.

Papers appear hours before press coverage β€” the fastest signal of what is happening at the AI research frontier.

Categories: cs.AI (Artificial Intelligence), cs.LG (Machine Learning), cs.CL (NLP), cs.CV (Computer Vision), cs.NE (Neural Computing), stat.ML.

All categories are fetched in ONE query (cat:A OR cat:B OR …), so the result is the most recent papers across the requested categories, grouped per category. A paper cross-listed to several requested categories appears under each of them, matching how a single-category query would return it.

Because one window is shared, a small category (cs.NE, stat.ML) can come back with fewer than limit papers even though arXiv holds more. Those categories are named in incomplete_categories β€” query a single category for full depth.

Args: params (ArxivLatestInput): - categories (List[str]): arXiv category codes - limit (int): Max papers per category (1–20)

Returns: str: JSON with categories, total_papers, distinct_papers, by_category dict, and β€” only when a category fell short β€” incomplete_categories plus an explanatory note. Each paper: id, title, abstract (400 chars), authors, published, category (primary), categories (all), url, pdf.

Inputs: params (required). Returns: note, categories, fetched_at, by_category, +3 more

hn_discussionA
Read-onlyIdempotent

Read the comment thread under a HackerNews story.

Where hn_top_stories and hn_search tell you what is being discussed, this tells you what is actually being argued β€” the counter-arguments, the practitioner caveats, the "we tried this in production" replies that carry the real signal. Algolia can search comment text but does not return thread structure, so this is the only way to see who replied to whom.

Get a story_id from hn_top_stories or hn_search first.

Comments are walked breadth-first, so the highest-ranked top-level comments come back first. Deleted and flagged comments are skipped. Popular threads run to several hundred comments and each one costs a request upstream, so both depth and total count are capped β€” check the 'truncated' flag to see whether the thread was cut short.

Args: params (HnDiscussionInput): - story_id (int): HackerNews item ID - max_depth (int): Reply nesting levels (1–4, default 2) - max_comments (int): Total comment budget (1–100, default 25) - text_chars (int): Per-comment text truncation (100–2000)

Returns: str: JSON with story{}, total_comments (as reported by HN), fetched_comments, truncated, comments[]. Each comment: id, by, posted, text, reply_count, replies[] (same shape, nested).

Inputs: params (required). Returns: story, comments, truncated, fetched_at, +3 more

hn_top_storiesA
Read-onlyIdempotent

Fetch stories from any of the six HackerNews front-page feeds.

Feeds: 'top' (frontpage), 'best' (highest voted), 'new' (latest), 'ask' (Ask HN β€” questions to the community), 'show' (Show HN β€” projects people are shipping), 'job' (YC company job posts).

'show' is the strongest signal for what practitioners are actually building; 'ask' for what they are stuck on. Upstream, 'ask' and 'job' hold only ~30 items, so a large limit may return fewer results.

Args: params (HnTopStoriesInput): - feed (str): 'top', 'best', 'new', 'ask', 'show', or 'job' - limit (int): Stories to return (1–30) - min_score (int): Minimum score filter (job posts score 1)

Returns: str: JSON with feed, count, stories[]. Each story: id, type, title, url, score, comments, by, posted, hn_link.

Inputs: params (required). Returns: feed, count, stories, fetched_at

lobsters_hotA
Read-onlyIdempotent

Fetch the hottest stories from Lobste.rs, a curated tech community.

Lobste.rs is smaller and more technically focused than HackerNews. Invitation-only membership ensures higher signal-to-noise ratio.

Args: params (LobstersHotInput): - limit (int): Stories to return (1–25) - tag_filter (Optional[str]): Tag substring filter (e.g. 'ai', 'ml')

Returns: str: JSON with count, stories[]. Each story: title, url, score, comments, tags, submitter, submitted_at, lobsters_url.

Inputs: params (required). Returns: count, stories, fetched_at, tag_filter

tech_signal_digestA
Read-onlyIdempotent

Aggregate tech & AI signals from all four sources in one call.

The primary tool for a comprehensive daily or weekly tech intelligence briefing. Combines HackerNews, arXiv, Lobste.rs and GitHub into one structured JSON digest. Use 'focus' to filter for a specific topic.

Args: params (TechSignalDigestInput): - focus (Optional[str]): Topic filter (e.g. 'MCP', 'agents') - hn_limit (int): HN stories (1–10) - arxiv_limit (int): arXiv papers (1–10) - lobsters_limit (int): Lobste.rs stories (1–10) - github_limit (int): GitHub repos (1–10)

Returns: str: JSON digest with generated_at, focus, degraded_sources[] and sources{hn, arxiv, lobsters, github}. Each source has label, count and its items list.

     A source that could not be reached still appears, with count 0
     and an 'error' describing why, and its key is listed in
     degraded_sources β€” treat its absence as unknown, not as zero.
     The GitHub section may carry 'incomplete_topics' when part of
     the topic sweep failed. Only if all four sources fail does this
     tool return a plain error string instead of JSON.

Inputs: params (required). Returns: focus, sources, generated_at, degraded_sources

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.1/5.0

    Scored across 8 tools

    Disambiguation5/5

    Each tool targets a distinct source and action: HN feed retrieval vs keyword search vs thread reading; arXiv latest vs search; Lobsters hot; GitHub trending; and an aggregate digest. Overlap is minimal, and the aggregate tool is clearly positioned as a convenience rather than a replacement.

    Naming Consistency4/5

    All names use snake_case and mostly follow a source_descriptor convention (hn_, arxiv_, lobsters_, github_). The suffixes vary between verbs and nouns/adjectives (search, discussion, latest, hot), so the pattern is not a strict verb_noun convention.

    Tool Count5/5

    Eight tools is well-scoped for a four-source signal aggregator: three for HN, two for arXiv, one each for Lobsters and GitHub, plus an aggregator. No tool feels redundant or excessive.

    Completeness4/5

    Core workflows are covered: discover recent items per source, search HN and arXiv, read HN discussions, and aggregate across sources. Minor gaps exist (e.g., no keyword search for Lobsters or GitHub beyond hot/trending), but agents can work around them via the digest focus filter.

    Maintenance

    ActivityActive
    ResponsivenessLimited data

    Related MCP Connectors

    Related MCP Servers