Observing ATProtocol

Firehose Access

Jetstream endpoint: wss://jetstream2.us-east.bsky.network/subscribe

Simplified JSON stream (~50 posts/sec on Bluesky). Easier than raw CAR files.

Quick Pulse

import asyncio import json import websockets from datetime import datetime, timezone

async def pulse(duration_seconds=15): uri = 'wss://jetstream2.us-east.bsky.network/subscribe' posts, likes, follows = 0, 0, 0 hashtags = {}

start = datetime.now(timezone.utc)

async with websockets.connect(uri) as ws:
    while (datetime.now(timezone.utc) - start).seconds < duration_seconds:
        try:
            msg = await asyncio.wait_for(ws.recv(), timeout=1)
            data = json.loads(msg)
            
            commit = data.get('commit', {})
            collection = commit.get('collection', '')
            
            if collection == 'app.bsky.feed.post':
                posts += 1
                # Extract hashtags
                text = commit.get('record', {}).get('text', '')
                for word in text.split():
                    if word.startswith('#'):
                        tag = word[1:].lower().rstrip('.,!?')
                        hashtags[tag] = hashtags.get(tag, 0) + 1
            elif collection == 'app.bsky.feed.like':
                likes += 1
            elif collection == 'app.bsky.graph.follow':
                follows += 1
        except asyncio.TimeoutError:
            continue

return {
    'duration': duration_seconds,
    'posts': posts,
    'likes': likes,
    'follows': follows,
    'posts_per_min': posts * 60 // duration_seconds,
    'top_hashtags': sorted(hashtags.items(), key=lambda x: -x[1])[:5]
}

Using tools/firehose.py

Quick network sample

uv run python -m tools.firehose sample 30

Network analysis

uv run python -m tools.firehose analyze 60

Monitoring Feeds

The Atmosphere Feed

Cameron's curated ATProtocol discussion feed:

feed_uri = 'at://did:plc:gfrmhdmjvxn2sjedzboeudef/app.bsky.feed.generator/the-atmosphere'

async with httpx.AsyncClient() as client: resp = await client.get( 'https://public.api.bsky.app/xrpc/app.bsky.feed.getFeed', params={'feed': feed_uri, 'limit': 20} ) posts = resp.json().get('feed', [])

Recording Observations

Record observation as a cognition thought

uv run python -m tools.cognition thought "Network pulse: {posts_per_min} posts/min, {likes_per_min} likes/min. Top tags: #atproto, #bluesky"

Network Statistics (typical)

• ~50 posts/sec (~3000/min)

• ~250 events/sec total

• ~65% of events are likes

• Engagement ratio (likes:posts) typically 4-6:1

• Higher engagement ratio = healthy interaction patterns

Best Practices

• Sample for at least 15 seconds for meaningful data

• Note time of day - activity varies

• Cultural events drive spikes (e.g., BBB26, sports)

• Record observations as network.comind.observation records

• Compare pulses over time to identify patterns