The open-source signal — and the market data hiding behind no API

Every week we scan the same two things: what open source is rewarding, and where the valuable data is that nobody has made easy to get. The first tells us what to build with. The second tells us what to build. This week both were loud.

What the star charts are actually rewarding

Scroll GitHub Trending right now and one theme swallows the rest: AI coding agents, and the "skills" packs that feed them. Almost every fast-riser is either an agent harness or a bundle of instructions you drop into one.

• xai-org/grok-build — a coding-agent harness and terminal UI — was the single biggest gainer of the week, roughly +3,300 stars.
• MoonshotAI/kimi-cli ("your next CLI agent") is right behind it.
• And a wave of skill packs — Nutlope/hallmark (an "anti-AI-slop design skill" for Claude Code, Cursor, and Codex), mattpocock/skills, ibelick/ui-skills — are shipping curated instructions straight out of their authors' .claude directories.

The other durable crowd-pleaser is back too: run something enormous on hardware you own. JustVugg/colibri claims a 744B-parameter model on a 25GB-RAM consumer box in pure C. Above it all, the evergreen giants still top the daily board — and trackers still crown OpenClaw, the local-first assistant that went from 9k to 210k+ stars this year, as the fastest-growing open-source project on record.

Why it matters to us: "skills" quietly became its own category. That's leverage. A well-scoped skill pack lets us hand a client's agent our exact process — brand voice, QA checks, deploy steps — instead of re-explaining it every time. We're already building internal ones. The teams treating agents as infrastructure, not novelty, are the ones shipping faster.

The platform with no API

The second signal is a data story. Ask "how do I get data out of Truth Social?" and every developer thread lands in the same place: there is no official API. No keys, no documented endpoints, no rate limits, no support.

It works anyway for one structural reason — Truth Social is a reskinned, non-federated fork of Mastodon. It inherited Mastodon's internal API surface, so every tool that touches it is calling undocumented endpoints, not a supported product. The community client, truthbrush, was even archived in April and re-forked when the surface shifted — which tells you exactly how brittle this is.

Why we actually care: it's a data feed, not a social network

Here's the part that makes it interesting to a firm that builds fintech and trading products. Truth Social's parent, Trump Media & Technology Group, trades publicly as DJT — and posts on the platform routinely move markets: individual equities, crypto, and DJT itself.

That reframes the whole thing. Truth Social isn't a social network you scrape for fun. It's a market-moving headline wire with no official API. Being first to a new post can matter in seconds — which is precisely why low-latency archives like stiles/trump-truth-social-archive exist, and why sentiment and quant shops pay for managed feeds from services like ScrapeCreators and SocialCrawl.

High-value data, guarded by a fragile, undocumented surface. That's not a bug in the opportunity — that is the opportunity.

What we tell clients

The same lesson runs through both halves of this week:

1. Own the pipe. When the data matters — and market-moving posts matter — you don't want to be one vendor's pricing change away from losing access. Build the connection, or pay someone whose only job is keeping it alive.
2. Expect breakage. Unofficial surfaces move. Archived clients get re-forked, endpoints shift overnight. Design for it: proxies, retries, monitoring, a fallback feed.
3. Move on the signal early. Agents-as-infrastructure and skills-as-leverage aren't hype cycles to wait out. The teams adopting them now are compounding a speed advantage the rest will spend next year catching up to.

We read the charts so our clients don't have to. Then we build the thing the chart was pointing at.