I dug through GitHub Trending and pulled the 5 repos that actually matter this week. Below each one: what it does, what I'd use it for, the real install command, and a link straight to the repo.
→PrimeIntellect shipped a coding agent that spawns recursive sub-agents and refines its own skills. $1B company, $130M raise last month. 15,880 stars, 12,476 of those from this week alone.
→Semantica built what they call an open source Palantir. Every AI decision becomes a queryable graph node with full W3C provenance. 7,450 stars.
→Google open-sourced 90+ agent skill packs for Claude Code, Codex, and Antigravity. One install command covers GKE, BigQuery, Gemini, and Ads. 18,172 stars.
→Deno's team built celld: self-hosted Durable Objects in Rust. Each object is its own SQLite DB replicated to your S3 bucket. No control plane. 3,532 stars.
🔒Drop your email for the full breakdown
All 5 repos with install commands, use cases, and what each one actually does under the hood. Plus the weekly builds I ship with these tools.
Join 3,000+ builders already getting this. No spam.
ALL 5 REPOS — WHAT THEY DO, HOW TO INSTALL
01
PrimeIntellect / prime-agent
A coding agent built on a persistent Python REPL instead of a chat window. Calling rlm(...) spawns real child agents that run in parallel and hand their results back programmatically, and /refine lets the agent rewrite its own supplemental prompts, memories, and skills based on what just happened in the session. Sessions run daemon-backed, so a long job keeps going after you close the terminal.
Stars
15,880 (+12,476 this week)
Language
TypeScript
Backing
$1B valuation, $130M Series A
curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh
Use it for: long-running research or refactor jobs you want to leave overnight · fan-out work where one agent farms 10 sub-agents at once · workflows you repeat weekly and want the agent to turn into a reusable skill.
Ingests your data into a knowledge graph where every AI decision becomes a node with full causal ancestry. You can trace any output back through the exact facts and rules that produced it, gate decisions behind policies, and export the whole chain as W3C PROV-O provenance. Self-hostable, MIT licensed, polyglot graph storage (RDF and LPG).
Stars
7,450 (+4,073 this week)
Language
Python
License
MIT
pip install semantica
Use it for: any agent making decisions a human will later be asked to justify (lending, claims, hiring, compliance) · replacing a vector store when the questions are relational rather than fuzzy · giving an internal AI tool an audit trail before legal asks for one.
90+ Agent Skills covering Google products, written by Google. Each one is procedural knowledge your agent loads on demand: how to authenticate to GCP, how to onboard a project, how to run GKE inference, how to query BigQuery, how to call the Gemini API. GKE alone has 25 skills. The install command lets you pick individual skills instead of taking all 90.
Stars
18,172 (+2,359 this week)
Skills
90+ across 9 categories
Works with
Claude Code, Codex, Antigravity
npx skills add google/skills
Use it for: stopping Claude from guessing gcloud syntax · deploying an agent to Cloud Run or GKE without reading the docs yourself · BigQuery analysis, Gemini API wiring, and Google Ads or Analytics pulls. This is the one I'd install today.
Cloudflare's Durable Objects, running on machines you own. Every object is its own SQLite database, continuously replicated to an S3-compatible or GCS bucket. Nodes coordinate through compare-and-swap on that bucket alone, so there is no control plane, no consensus service, and no membership protocol to operate. An idle object costs close to nothing, and apps shard by construction. Rust, from Ryan Dahl's team.
Stars
3,532 (+1,783 this week)
Language
Rust
License
Apache-2.0
curl -fsSL https://celld.dev/install.sh | sh
Use it for: per-user or per-session state (chat threads, agent memory, game rooms) where one shared Postgres becomes the bottleneck · getting off Cloudflare while keeping the Workers programming model · edge-style apps on your own hardware. Docker image is published too: ghcr.io/denoland/celld.
Parses a multi-language codebase with Tree-sitter, builds a Memgraph knowledge graph of its actual structure, then lets you query and edit that code in plain English through AST-surgical patches. A runtime tracer can run your test suite and merge the calls that really happened into the graph as CALLS edges. It ships as an MCP server, so Claude Code talks to it directly.
Use it for: monorepos too big for an agent to hold in context · impact analysis before a refactor (what actually calls this) · onboarding to a codebase you inherited. Needs Docker for Memgraph, plus cmake and ripgrep. Then run cgr daemon up.
