Browsing Bluesky

Access Bluesky content through public APIs and real-time firehose. Supports optional authentication for personalized feeds. Includes account analysis for categorization.

Implementation

Add skill directory to path and import:

import sys
sys.path.insert(0, '/path/to/skills/browsing-bluesky')  # or use .claude/skills symlink path
from browsing_bluesky import (
    # Core browsing
    search_posts, get_user_posts, get_profile, get_feed_posts, sample_firehose,
    get_thread, get_quotes, get_likes, get_reposts,
    get_followers, get_following, search_users,
    # Trending
    get_trending, get_trending_topics,
    # Account analysis
    get_all_following, get_all_followers, extract_post_text,
    extract_keywords, analyze_account, analyze_accounts,
    # Authentication utilities
    is_authenticated, get_authenticated_user, authenticated_identity,
    clear_session
)

Authentication (Optional)

Authentication enables personalized feeds (like Paper Skygest) that require knowing who's asking.

Setup

1. Create an app password at Bluesky: Settings → Privacy and Security → App Passwords
2. Set environment variables:

   export BSKY_HANDLE="yourhandle.bsky.social"
   export BSKY_APP_PASSWORD="xxxx-xxxx-xxxx-xxxx"
   

Which account reads

Two credential pairs are recognised, and the prefix is the only difference:

Pair	Identity
MUNINN_BSKY_HANDLE / MUNINN_BSKY_APP_PASSWORD	muninn — preferred
BSKY_HANDLE / BSKY_APP_PASSWORD	owner — used when no Muninn pair is set

BSKY_IDENTITY=muninn|owner picks one outright, and a pair named that way is never substituted for: asking for muninn with no Muninn pair set reads as public rather than quietly reading as the owner. An unrecognised value raises.

This matters because a booted container holds both pairs. Until 2026-08-28 the unprefixed one won, so every authenticated read — the following timeline, and the mutes, blocks and labelers that shape all of them — came back as the account owner with nothing in the output saying so, from a skill documented as read-only that was nonetheless holding an app password carrying write scope on his account. On a machine with only the documented BSKY_* pair, nothing changes.

Behavior

• Transparent: All functions work identically with or without credentials
• Automatic: Auth headers are added opportunistically when credentials exist
• Graceful: Failed auth silently falls back to public access. A misspelled BSKY_IDENTITY is configuration, not auth failure, and raises instead
• Secure: Tokens cached in memory only, never logged or persisted

Check Auth Status

if is_authenticated():
    print(f"Logged in as: {get_authenticated_user()}")
else:
    print("Using public access")

# Which pair answered — a handle alone does not say whether it was chosen
# or simply the only one set.
authenticated_identity()   # {'identity': 'muninn', 'handle': ..., 'did': ...}

# Clear session if needed (e.g., switching accounts or identities)
clear_session()

Research Workflows

Investigate a Topic

Use search_posts() with query syntax matching bsky.app advanced search:

• Basic terms: event sourcing
• Exact phrases: "event sourcing"
• User filter: from:acairns.co.uk or use author= param
• Date filter: since:2025-01-01 or use since= param
• Hashtags, mentions, domain links: #python mentions:user domain:github.com

Combine query syntax with function params for complex searches.

Monitor a User

1. Fetch profile with get_profile(handle) for context (bio, follower count, post count)
2. Get recent posts with get_user_posts(handle, limit=N)
3. For topic-specific user content, use search_posts(query, author=handle)

Discover What's Trending

Recommended workflow — trending API first, firehose for deep dives:

1. Quick scan with trending topics (~500 tokens)

topics = get_trending_topics(limit=10)
# Returns: {topics: [{topic, display_name, description, link}, ...],
#           suggested: [...]}

2. Rich trends with post counts and actors

trends = get_trending(limit=10)
for t in trends:
    print(f"{t['display_name']} — {t['post_count']} posts ({t['status']})")
# Each trend includes: topic, display_name, link, started_at,
#   post_count, status, category, actors

3. Targeted exploration of selected trends

posts = search_posts(trend["topic"], limit=25)

4. Optional: Firehose for velocity monitoring or long-tail discovery

Prerequisites: Install Node.js dependencies once per session:

cd /home/claude && npm install ws https-proxy-agent 2>/dev/null

data = sample_firehose(duration=30)  # Full firehose sample
data = sample_firehose(duration=20, filter="python")  # Filtered sample

Returns dict with keys:

• window: {startTime, endTime, durationSeconds} — sampling time range
• stats: {totalReceived, totalPosts, postsPerSecond, filter, languages} — volume metrics and language breakdown
• topWords: [[word, count], ...] — top 50 words (count >= 3)
• topPhrases: [[bigram, count], ...] — top 30 bigrams (count >= 2)
• topTrigrams: [[trigram, count], ...] — top 20 trigrams (count >= 2)
• entities: [[entity, count], ...] — top 25 handles/hashtags (count >= 2)
• samplePosts: [{text, altTexts, hasImages}, ...] — first 50 matching posts

Read Feeds and Lists

get_feed_posts() accepts:

• List URLs: https://bsky.app/profile/austegard.com/lists/3lankcdrlip2f
• Feed URLs: https://bsky.app/profile/did:plc:xxx/feed/feedname
• AT-URIs: at://did:plc:xxx/app.bsky.graph.list/xyz

The function extracts the AT-URI from URLs automatically.

Explore a Thread

Fetch full thread context for a post with parents and replies:

thread = get_thread("https://bsky.app/profile/user/post/xyz", depth=10)
# Returns: {post: {...}, parent: {...}, replies: [...]}

Find Quote Posts

Discover posts that quote a specific post:

quotes = get_quotes("https://bsky.app/profile/user/post/xyz")
for q in quotes:
    print(f"@{q['author_handle']}: {q['text'][:80]}")

Analyze Engagement

Get users who engaged with a post:

likes = get_likes(post_url)
reposts = get_reposts(post_url)

# Accepts both URLs and AT-URIs
likes = get_likes("at://did:plc:.../app.bsky.feed.post/...")

Read Embed Images

Every parsed post carries an images field — a list of {alt, url, transcription} dicts, one per embed image. The legacy image_alts: list[str] field is preserved (non-empty alts only).

When alt text is missing and the image content matters, opt in to model transcription via the transcribe parameter on any post-fetch function (get_user_posts, search_posts, get_feed_posts, get_thread, get_quotes):

# Routine/bulk work (zeitgeist, inbox review, news scans) —
# gemini-2.5-flash-lite is the recommended default. Cheapest production
# model anywhere ($0.10/$0.40 per 1M tokens), ~95% accuracy on dense
# screenshots in May 2026 benchmarks:
posts = get_user_posts("ayourtch.bsky.social", limit=40, transcribe="gemini-lite")

# Token-perfect transcription, still cheap:
posts = get_user_posts(..., transcribe="gemini-flash")

# Frontier model with thinking_level=minimal — for cases where the image
# content needs reasoning, not just transcription:
posts = get_user_posts(..., transcribe="gemini-3.5-flash")

# Anthropic single-vendor option (note: empirically weaker prompt-following
# than Gemini on dense transcription — Haiku tends to summarize rather
# than transcribe):
posts = get_user_posts(..., transcribe="haiku")

# Interactive sessions where image is part of the active task and you want
# conversation context to inform interpretation (only available on Anthropic):
thread = get_thread(post_url, transcribe="opus")

# Default (no transcription) — current behavior preserved:
posts = get_user_posts("ayourtch.bsky.social", limit=40)

Policy is invariant across all callers: images with non-empty alt text are never transcribed (the author already described the image; trust it). Only images with missing or empty alt are sent to the model. Network or API failures leave transcription as None; callers degrade silently.

Cost/quality empirics (May 2026, n=3 dense terminal screenshots, single run each — sample size is small, treat as directional):

Alias	Latency	$/image	Chord-token recall
gemini-lite	~8s	~$0.001	95%
gemini-flash	~10s	~$0.003	100%
gemini-3.5-flash	~10s	~$0.014	100%
haiku	~7s	~$0.008	18% (summarizes)
opus	~20s	~$0.12	91%

The haiku and opus rows were measured on Haiku 4.5 and Opus 4.7. The aliases now resolve to claude-haiku-5-5 and claude-opus-5-5; those rows have not been re-measured.

Requires either ANTHROPIC_API_KEY (or API_KEY in /mnt/project/claude.env) for the haiku / opus aliases, or CF AI Gateway credentials in /mnt/project/proxy.env for the gemini-* aliases. Transcription only fires when the parameter is set, so callers without the relevant credentials can simply pick a different alias or leave the feature off.

Explore Social Graph

Navigate follower/following relationships:

followers = get_followers("handle.bsky.social")
following = get_following("handle.bsky.social")

# Returns list of actor dicts with handle, display_name, did, description, etc.

Find Users

Search for users by name, handle, or bio:

users = search_users("machine learning researcher")
for u in users:
    print(f"{u['display_name']} (@{u['handle']}): {u['description'][:100]}")

API Endpoint Notes

• Public AppView: https://api.bsky.app/xrpc/ for unauthenticated reads
• PDS: https://bsky.social/xrpc/ for authenticated requests
• Trending: app.bsky.unspecced.getTrends (rich) and app.bsky.unspecced.getTrendingTopics (lightweight)
• Firehose: wss://jetstream1.us-east.bsky.network/subscribe
• Endpoint routing is automatic - authenticated requests go to PDS, public requests go to AppView
• Rate limits exist but are generous for read operations

Return Format

All API functions return structured dicts with:

• uri: AT protocol identifier
• text: Post content
• created_at: ISO timestamp
• author_handle: User handle
• author_name: Display name
• likes, reposts, replies: Engagement counts
• links: Full URLs extracted from post facets (post text truncates URLs with "...")
• image_alts: Alt text from embedded images
• url: Direct link to post on bsky.app

Profile function returns: handle, display_name, description, followers, following, posts, did

Account Analysis

Analyze accounts for categorization by topic. Fetches profile and posts, extracts keywords, and returns structured data for Claude to categorize.

Analyze a User's Network

# Analyze accounts you follow
results = analyze_accounts(following="yourhandle.bsky.social", limit=50)

# Analyze your followers
results = analyze_accounts(followers="yourhandle.bsky.social", limit=50)

# Analyze specific handles
results = analyze_accounts(handles=["user1.bsky.social", "user2.bsky.social"])

Single Account Analysis

analysis = analyze_account("user.bsky.social")
# Returns: {handle, display_name, description, keywords, post_count, followers, following}

Keyword Extraction Options

Stopwords parameter filters domain-specific noise:

• "en": English (general purpose, default)
• "ai": AI/ML domain (filters tech boilerplate)
• "ls": Life Sciences (filters research methodology)

results = analyze_accounts(following="handle", stopwords="ai")

Requires: extracting-keywords skill with YAKE venv for keyword extraction.

Filtering Accounts

results = analyze_accounts(
    following="handle",
    exclude_patterns=["bot", "spam", "promo"]  # Skip accounts matching these
)

Paginated Following/Followers

For large account lists beyond the 100 limit of get_following/get_followers:

all_following = get_all_following("handle", limit=500)  # Handles pagination
all_followers = get_all_followers("handle", limit=500)

Account Analysis Output

Each analyzed account returns:

{
    "handle": "user.bsky.social",
    "display_name": "User Name",
    "description": "Bio text here",
    "keywords": ["keyword1", "keyword2", "keyword3"],
    "post_count": 20,
    "followers": 1234,
    "following": 567
}

Claude uses bio + keywords to categorize accounts by topic without hardcoded rules