Bluesky Protocol Services: Jetstream v2 with Network Replay
Bluesky launched Bluesky Protocol Services, a new brand and website for the public infrastructure it runs on the AT Protocol network. The big news is Jetstream v2, which adds Network Replay — a way to get historical network data over HTTP and then cut over to the live WebSocket stream without a gap.
Jetstream v2: Network Replay
Jetstream is a service that lets you subscribe to a slice of the AT Protocol network as plain JSON over WebSocket. Previously, you couldn't get historical records from Jetstream; you had to backfill repos yourself. Jetstream v2 fixes that with a new Network Replay feature.
Here's how it works:
- POST your filters to
planSnapshotto get a plan for historical data. - Download the sealed segments over plain HTTP.
- Connect to the live WebSocket once at the tip to continue in real-time.
The replay is stateless on the server: no per-consumer cursor, no subscription registration, nothing to stage client-side. You can also just snapshot the network (point-in-time copy) over HTTP only, using listSegments and getSegment, with no live tail.
This enables server-side analysis over historical data without backfilling locally. Use cases include spinning up an App, running analysis over a month of posts, or recovering from downtime.
Because serving archives is bandwidth-intensive, Bluesky is now requiring an API token for archive requests. The live tail remains open and unauthenticated. No plans to add auth to the live stream.
The v2 instances are live now:
wss://jetstream.us-west.bsky.network
wss://jetstream.us-east.bsky.networkThe existing v1 instances keep running unchanged, and the live tail behaves identically, so no rush to migrate.
New Jetstream SDKs
They also released official Jetstream SDKs for TypeScript and Go. Here's the TypeScript example from the docs:
import { Jetstream } from '@bsky/jetstream'
import { app } from '@bsky/sdk/lexicons'
const js = new Jetstream('https://jetstream.us-east.bsky.network')
for await (const evt of js.live({ collections: [app.bsky.feed.post] })) {
if (evt.kind === 'commit' && evt.commit.operation === 'create') {
console.log(evt.commit.collection, evt.commit.record.text)
}
}
The TypeScript SDK is on npm; the Go SDK is part of the Jetstream project.
Bluesky TypeScript SDK Rebased on lex
The Bluesky TypeScript SDK is rebuilt on @atproto/lex, removing legacy code paths. It's fully typed end-to-end, from protocol layer up through app.bsky records. All new docs examples use it. The old @atproto/api still works, but the new SDK eliminates LLM recommendations of deprecated versions.
Updates to endpoints.bsky.app
The HTTP reference at endpoints.bsky.app has been updated. The old docs.bsky.app is replaced by the new Protocol Services site.
All infrastructure is open source and self-hostable.
📖 Read the full source: HN AI Agents
👀 See Also

Local LLM Benchmark: Backend Generation by Function Calling – GLM, Qwen, DeepSeek Compared
A rigorous benchmark of local and frontier LLMs for backend code generation via function calling, with scoring rubric. Key findings: qwen3.5-35b-a3b matches gpt-5.4 on DB/API design, and dense Qwen 27B beats 397B MoE. Frontier models dropped due to cost.

Claude Code v2.1.133: worktree.baseRef revert, sandbox paths, proxy fix for MCP OAuth
Anthropic releases v2.1.133 of Claude Code CLI with a new worktree.baseRef setting defaulting to fresh (branch from origin/default), sandbox.bwrapPath and sandbox.socatPath for custom bubblewrap/socat binaries, proxy/mTLS fix for MCP OAuth flow, and several bug fixes.

Claude Status Update: Elevated Error Rates for Opus 4.6 and Sonnet 4.6
An official Claude system status update reports elevated error rates for Opus 4.6 and Sonnet 4.6 models, with an incident timestamped 2026-03-31T21:10:28.000Z. The automatic post directs users to check resolution status and community performance reports.

Uber burns 2026 AI budget in 4 months on Claude Code — $500–$2k per engineer monthly
Uber spent its entire 2026 AI budget by April on Claude Code and Cursor. Monthly API costs hit $500–$2,000 per engineer. 95% of engineers use AI tools monthly; 70% of committed code is AI-generated.