Bring-your-own-key · No proxying · Your provider, your bill
Today's snapshot
What changed overnight
Major SDKs are settling on the same shapes for tool calls, so switching models should mean less glue code and fewer surprises. The eval harness many teams share quietly re-weighted long, multi-step tasks—old leaderboard scores aren't apples-to-apples. Quiet on arXiv, but two repo threads are worth a skim before standup.
01
Tool calls, one dialect
Frameworks are converging on compatible tool JSON; you write adapters once, not per provider.
02
Re-baseline your evals
A widely used harness changed how it scores agentic runs—rerun benchmarks before you trust last week.
03
Two diffs to skim
Streaming rerank for retrieval quality, plus smarter eviction when long-session memory fills up.
One brief. Every source you actually trust.
Any OpenAI-compatible key. Set the base URL and model you trust. Lighthouse never sees the key — it stays on your device.
Each brief is one declarative file: feeds, arXiv categories, GitHub queries, forum keywords. Add a file, get a new brief.
Email, Slack, Discord, Telegram, Notion. Each brief is formatted for its destination — Slack mrkdwn, Telegram HTML, styled email. Skip channels you don’t use.
Embeds every document, finds near-duplicates, and clusters the noisy ten posts into the one thing they’re really about.
A Kestra-orchestrated pipeline. Every run, every retry, every artifact is a click away. Debugging looks like a normal job.
No middleman tokens, no proxying. Your AI costs go to your provider. Free to use, free to fork, fair to all.
Each morning, Lighthouse fans out to your sources, dedupes the noise, classifies what matters, and writes the brief — all orchestrated by Kestra. You wake up to one Markdown page, not 47 tabs.
# topics/agentic-eng.yaml
id: agentic-eng
name: Agentic Engineering
schedule: "0 8 * * *"
sources:
rss:
- https://simonwillison.net/atom.xml
- https://anthropic.com/news/feed.xml
arxiv_categories: [cs.AI, cs.CL]
github_queries:
- language:python topic:agents
hn_keywords: [llm, agent, mcp]
reddit_subs: [LocalLLaMA]
deliver:
email: you@studio.com
slack: $SLACK_WEBHOOK
discord: $DISCORD_WEBHOOK
telegram: $TELEGRAM_CHAT_ID
notion: $NOTION_PAGE_IDTrack 8 forums and 12 RSS feeds without opening a single tab.
A daily arXiv + GitHub digest, scoped to your sub-field.
Every conversation about your tool, on one Markdown page.
One brief in the morning — not a Slack channel from hell.
AI coding agents, orchestration frameworks, MCP, evals, and the practical engineering stack around production LLM applications.
37 sources tracked
Solana ecosystem development (Anchor, Pinocchio, SVM internals) plus zero-knowledge cryptography research, audit reports, and ZK rollup news.
36 sources tracked
Bootstrapped + small-team SaaS — distribution, pricing, growth, building in public, and the operational stack solo founders use.
26 sources tracked
Modern data stack — orchestration, lakehouse formats, query engines, and the AI-native data layer (vector DBs, embedding pipelines, RAG infra).
36 sources tracked
Make a free account, paste your AI key, pick a topic, choose a channel. Your first brief takes about a minute to set up.