1. cathrynlavery/diagram-design — Diagrams from your agent you can actually publish
Ask a model for an architecture diagram and you get a pile of rounded boxes that match nothing else on your site. Settle for it and the diagram was pointless; don't, and you're in Figma for half an hour. The author says she got stuck at exactly this step every time she wrote a post, so she wrote down what makes a diagram good and turned it into rules: fixed layouts, fixed palette, the accent color reserved for the one or two things a reader should notice first, target density 4/10 — every node has to earn its place.
Installed, it's a skill for Claude Code, Codex, and Pi, with dedicated templates for architecture sketches, flowcharts, pyramids, and more (the README says 27 visual types, the repo description says 29 — check the actual directory). Output is self-contained HTML + SVG: no build step, no external image dependency, drop it straight into a site. Already sitting on old draw.io or Mermaid sources? It'll redraw them at a size and level of detail you pick.
MIT, roughly +14,700 stars this week, comfortably first on the weekly board. Worth installing if you write technical posts, product docs, or internal proposals — assuming you already live in one of those agents. If your diagrams end up back in Figma for polish, this doesn't fit that pipeline.
2. semantica-agi/semantica — Making enterprise AI explain itself
Inside a company, getting an answer out of a model usually isn't the hard part. Explaining afterwards which data it rested on and which steps it took is. Vector retrieval can't give you that chain — it only tells you which passages looked similar. Semantica takes another route: ingest enterprise data, extract it into a context graph and a knowledge graph, run analytics and causal reasoning on the graph, and hang queryable decision provenance off every conclusion. The project calls itself "the open source Palantir for AI agents," supports both RDF and LPG graph models, follows W3C standards, and lets you manage your own ontology.
Python, MIT, self-hostable, roughly +5,300 stars this week. Finance, healthcare, public services — anywhere you have to justify a model's reasoning to a compliance function — is the obvious home turf. The price is that your team needs real knowledge-modeling and ontology skills. If you just want an internal Q&A bot, this is a lot of machinery to keep running.
3. PrimeIntellect-ai/prime-agent — A coding agent that builds up its own experience
Most agents keep everything in the conversation context, so they go blank when the window fills and yesterday's hard-won working style has to be taught all over again. Prime Intellect attacks that with two abstractions. The first is the RLM, or Recursive Language Model: context becomes variables, subagents become function calls, and it all runs inside a persistent Python REPL where reading files, running commands, and managing context are just code — call rlm(...) and you've spawned a real child agent to work in parallel. The second is the Continual Harness (paper here), which keeps supplemental prompts, memories, skill descriptions, and reusable subagent specs as durable state, refined through small evidence-backed edits that stay local to the session by default.
Put plainly: useful context and working patterns that outlive a single chat window. TypeScript, MIT, roughly +8,500 stars this week. Worth trying on refactors, migrations, or research that runs for hours or days — provided you're fine with an agent orchestrating everything in code rather than prose. If what you want is autocomplete, an editor plugin is less work.
4. NVIDIA-NeMo/Switchyard — Point your coding agent at any model
Tools like Claude Code and Codex only speak their vendor's API, so the moment you want an open-source model or your own inference stack behind them, you're stuck at step one. The NVIDIA NeMo team wrote a Rust proxy that sits in the middle and translates: OpenAI Chat, Anthropic Messages, and OpenAI Responses convert between each other, the agent keeps speaking its native API, and the actual work goes to vLLM, NVIDIA NIM, Ollama, or any OpenAI-compatible endpoint. While it's there, it can split traffic across several models by whatever algorithm you configure, with Prometheus metrics covering latency, tokens, and routing overhead.
Apache-2.0, roughly +1,300 stars this week. If you want self-hosted models driving off-the-shelf agents, or you're benchmarking models and costs, spin it up and play. But the README carries its own warning: pre-alpha, not for production, with the API and algorithms set to change significantly before v1.0. Putting it in a critical path today is asking for trouble.
5. vitali87/code-graph-rag — Your whole monorepo as one graph
Chunk your code into a vector store and what comes back merely looks similar; it can't tell you who calls this function or what breaks when you delete it. This project parses a multi-language codebase with Tree-sitter into functions, classes, methods, modules, and the relationships between them, stores it in Memgraph, and then lets you query, edit, and optimize in plain language: pull real source by name or intent, patch surgically through the AST with a diff preview first, walk call edges from entry points to find dead code.
The neat recent addition is runtime tracing — run your test suite and the calls that actually happened get merged back into the graph, filling in the dispatch through interfaces, virtual methods, reflection, and framework routing that static analysis never sees, flagged where it was missed. Python, MIT, roughly +1,800 stars this week. It pays off most in a multi-language monorepo you're refactoring, documenting, or de-cluttering. The cost is running Memgraph yourself; on a small repo, editor search will beat it every time.
6. cactus-compute/needle — 14MB of tool calling, small enough for a watch
On phones, watches, speakers, and robots, a lot of the job is just "understand one sentence, then call the right function." Making a cloud round trip for that loses on latency, privacy, and offline use all at once. Needle 2 goes straight at that case: 45M parameters, and by the project's own account a single 14MB binary that runs a full session in about 28MB of RAM, quantized to 2 bits with the team's CQ2 scheme (weights on Hugging Face, architecture in arXiv:2607.18363).
A few engineering choices matter more here than the parameter count. Output is constrained by a byte-level grammar compiled from your own schemas, so what comes back is structured JSON. Every response carries a calibrated confidence score: act above your threshold, escalate below it. And when the tool catalogue gets long, a retrieval head renders only the five most relevant per turn. MIT, roughly +2,500 stars this week. Worth a serious look for on-device, wearable, robotics, and offline work — just don't ask it for prose or hard reasoning, because it isn't a general chat model. The comparisons against FunctionGemma 270M and LFM2.5 230M are the team's own numbers, so re-run them before you ship.
7. macro-inc/macro — One workspace, one memory, shared with your agents
Slack, Linear, Notion, and HubSpot are all fine on their own. The problem is that gluing them together with MCP and Zapier scatters the state of the company across several SaaS products — people context-switch all day, and agents can't see any of it whole. Macro's answer is to build one thing: email, chat, docs, tasks, calls, CRM, and agents in a single workspace, everything @-linked, with cross-references between a doc and a task or a channel message and an email stored natively as a bidirectional graph, so humans and agents read the same context. The team, in New York and Toronto, dogfooded it for two years at ~15 people before opening it up.
SolidJS + Rust, roughly +2,400 stars this week. Realistically it's for startups of a few to a few dozen people willing to swap out their whole stack; if you're deep into existing tools, price the migration honestly. One thing to check first: the license is AGPL-3.0, which is a hard constraint for anyone planning closed-source derivatives or a SaaS built on top.
8. TencentCloud/TencentDB-Agent-Memory — A memory hub for the whole team's agents
On a team, every person's agent remembers separately, so the trap a colleague hit last week and the conventions everyone agreed on get explained again with each new person and each new tool. Tencent Cloud's project pulls memory out of the agent into a standalone service, distilling conversations, docs, and code into four reusable assets — Chat Memory, Skill, LLM-Wiki, and Code-Graph — that several agents and frameworks can share off one memory server, with governance over what gets shared.
Deployment is three pieces (memory-core, memory-hub, proxy) started by a single command, with a local panel. TypeScript, roughly +4,000 stars this week. It suits teams that want to accumulate team knowledge rather than personal chat logs; solo users will find it heavy, since Node 22.16+ plus three self-hosted services is a real setup. One licensing wrinkle: the README carries an MIT badge, but GitHub doesn't recognize a standard license for the repo, so read the LICENSE file yourself before commercial use.
The week in review
Not one of these eight is a new foundation model, and that's the loudest signal of the week. The heat has moved to the middle layer: memory for agents (semantica, TencentDB Agent Memory, macro's shared memory), skills (diagram-design's rules for drawing), routing (Switchyard), and code understanding (code-graph-rag). The top three read like two solutions to one problem — diagram-design has a human encode hard-won judgment as rules for the agent, prime-agent has the agent persist its own working patterns — and both concede the same premise: the models are smart enough already; the hard part is consistency.
The other thread runs downward. Needle 2 squeezes tool calling into 14MB, which means a model small enough for a watch is no longer a demo, it's a real option for on-device agents.
So if you already use a coding agent, the better use of this week is giving it skills and memory rather than swapping models again. If you're deploying inside a company, the governable, traceable memory layer that semantica and TencentDB Agent Memory represent is probably unavoidable this year. And one last thing: trending isn't ready. Switchyard says so itself, macro is AGPL-3.0 — check the license and the maturity before you commit.
via: GitHub Trending weekly, each project's GitHub repository and README; star counts and weekly deltas as of 2026-08-16, verified 2026-08-16