GitHub Trending Top 10: Diagram skills, accountable graphs, and tiny-device agents (Aug. 9–16)
This week's top ten show open-source projects turning diagrams, context, model routing, code understanding, and team workflows into explicit, reusable layers, with a star-or-skip verdict for every repository.
The Aug. 9–16 chart is less about another chatbot than about the layers around one. It includes a diagram skill, two kinds of knowledge graph, a long-running coding agent, a model router, a tiny on-device model, a unified work workspace, a low-code platform, and a skill distribution pack. The useful question is what boundary each project makes visible—and whether that boundary is worth keeping in your own toolbox. 1
Coverage: Aug. 9, 2026 at 20:00 through Aug. 16, 2026 at 20:00 Eastern Time. The rank and weekly star-growth figures below are the GitHub Trending weekly-page snapshot captured at the publication boundary. Repository API metadata was captured separately, so its current star and activity fields can differ by a small amount. 1
The chart at a glance
| Rank | Repository | Language | Stars on weekly page | Stars this week | License |
|---|---|---|---|---|---|
| 1 | cathrynlavery/diagram-design 2 | HTML | 19,529 | 15,600 1 | MIT |
| 2 | semantica-agi/semantica 3 | Python | 8,184 | 5,284 1 | MIT |
| 3 | PrimeIntellect-ai/prime-agent 4 | TypeScript | 16,572 | 6,435 1 | MIT |
| 4 | NVIDIA-NeMo/Switchyard 5 | Rust | 1,687 | 1,435 1 | Apache-2.0 |
| 5 | megadose/holehe 6 | Python | 13,293 | 1,287 1 | GPL-3.0 |
| 6 | cactus-compute/needle 7 | Python | 6,574 | 2,950 1 | MIT |
| 7 | macro-inc/macro 8 | Rust | 3,407 | 2,588 1 | AGPL-3.0 |
| 8 | vitali87/code-graph-rag 9 | Python | 4,429 | 1,686 1 | MIT |
| 9 | ToolJet/ToolJet 10 | JavaScript | 40,033 | 1,518 1 | AGPL-3.0 |
| 10 | addyosmani/agent-skills 11 | JavaScript | 87,761 | 2,882 1 | MIT |
1. cathrynlavery/diagram-design
Most AI-generated diagrams fail before the content is wrong: they are generic rounded boxes with no visual hierarchy.
diagram-design turns diagram-making into an installable skill for Claude Code, Codex, and Pi. It ships 27 visual types—flowcharts, timelines, data flows, security matrices, loops, and more—as self-contained HTML and SVG. Static output needs no build step, JavaScript, or external image dependency. 12The interesting implementation choice is the separation of semantic pattern from visual layout. A queue, policy trace, or trust boundary can use a suitable existing type instead of creating another template. Onboarding reads a website, maps its colors and fonts to semantic roles, checks contrast, and stores a reusable style guide. That is closer to a small design system than to a prompt collection. The repository also provides a gallery and client-specific installation paths. 12
The trade-off is scope: this produces polished explanatory assets, not a collaborative design editor or a general charting library. The API snapshot shows 23 open issues, 17 contributor accounts, a last push on Aug. 14, and MIT licensing. The README documents installation and examples but does not present a separate CI or automated-test setup. 213
Verdict: star it if your team writes technical docs, proposals, or architecture notes. The reusable pattern is a constrained visual vocabulary with brand tokens; skip it if you need interactive design tooling.
2. semantica-agi/semantica
Semantica addresses a harder failure mode than weak retrieval: an agent can quote the right document and still be unable to explain how a decision followed from it. The project builds a context and knowledge graph with decision records, causal links, conflict detection, provenance, ontology management, and graph analytics. It is aimed at systems where an audit trail matters as much as an answer. 14
The stack is unusually broad but coherent. A Python pipeline ingests files, databases, streams, Git, email, and enterprise sources; extracts entities, relations, and events; then writes to RDF or labeled-property graphs. SHACL, OWL, W3C PROV-O, Rete, Datalog, and SPARQL provide governance, lineage, and deterministic reasoning. Oxigraph, Neo4j, FalkorDB, Apache AGE, Neptune, and vector stores are presented as swappable backends. 14
Its difference from a vector database is the data model: decisions and their reasons become queryable objects, while contradictory facts can be flagged instead of silently blended. The cost is operational weight and a large surface to validate. The API shows 77 open issues, 44 contributor accounts, a last push on Aug. 16, and MIT licensing. The README has architecture, installation, and module documentation; it does not spell out a separate CI/test system. 315
Verdict: star it if you build governed agent systems or need explainable context. Skip it as a drop-in RAG replacement; the graph and policy work are the product.
3. PrimeIntellect-ai/prime-agent
Prime Agent is built for work that should survive the end of a chat window. Its Recursive Language Model treats context as variables and recursive subagents as programmatic calls inside a persistent REPL. A continual harness stores prompts, memories, skills, and subagent specifications as durable, reviewable state. Daemon-backed sessions, goals, heartbeats, compaction, and agent-to-agent messaging extend the same idea across long tasks. 16
The implementation centers on a persistent IPython control environment with daemon, worker, and kernel processes; GitHub labels TypeScript as the primary language. The agent can run shell commands, edit files, call tools, spawn child agents, and refine supplemental harness state. That makes it a runtime for ongoing work rather than a chat wrapper with a bigger prompt. The README includes a quickstart, CLI reference, RLM programming model, JSON/RPC modes, and long-running-agent documentation. 16
The boundary that matters is trust. Prime Agent explicitly says model-generated Python and project commands run with the user's permissions, and its process isolation is not a security sandbox. The API snapshot shows 73 open issues, 214 contributor accounts across three contributor pages, a last push on Aug. 16, and MIT licensing. The documentation is substantial, while the README does not give a separate CI/test inventory. 4171819
Verdict: star it if you are evaluating persistent coding-agent runtimes. Keep the warning beside the star: use a disposable worktree or external sandbox for untrusted code.
4. NVIDIA-NeMo/Switchyard
Switchyard puts a routing layer between an agent and the model endpoint. Its Rust proxy and library translate OpenAI Chat, OpenAI Responses, and Anthropic Messages formats, then route traffic to vLLM, NVIDIA NIM, Ollama, or other compatible backends. It records request, error, latency, token, and routing-overhead metrics through Prometheus. 20
The architecture is useful because the client keeps its native protocol while the gateway chooses the backend. Routing can be random for an A/B split, classifier-driven when request content determines the tier, signal-driven when tool results and errors already provide evidence, or custom when the algorithm belongs in the host application. The library path deliberately does not own the HTTP stack, so it can slot into an existing gateway or agent runtime. 20
That is a cleaner separation than hard-coding provider logic into every agent, but Switchyard is candidly pre-alpha. The API shows 105 open issues, 23 contributor accounts, a last push on Aug. 16, and Apache-2.0 licensing. The README provides launcher, server, library, routing, and core-concept documentation; it warns that the API and algorithms may change significantly and the software is not for production use. 521
Verdict: star it as a model-gateway design reference or an experiment harness. Skip it for a production dependency until the pre-alpha boundary moves.
5. megadose/holehe
holehe answers a narrow OSINT question: whether an email address appears to be attached to accounts on a large set of services. It checks more than 120 sites through registration or password-recovery flows, returns a standard structured result, and can be used from the CLI or embedded in Python. That makes it a collection of service-specific probes rather than a general identity-resolution system. 22The implementation is simple and revealing. Each site is a Python module that reports whether an account appears to exist, whether the request was rate-limited, and whether a partially masked recovery address or phone number was returned. The standard output lets callers compose the checks without parsing terminal text. The README gives installation paths, CLI and Python examples, Docker usage, a Maltego transform, and a module table. 22
Its differentiation is breadth and a low integration barrier; its weakness is that the signal comes from external service behavior that can change, rate-limit, or produce ambiguous results. The API snapshot shows 104 open issues, 25 contributor accounts, no push since Sept. 10, 2024, and GPL-3.0 licensing. The README describes usage and examples but does not present a separate CI/test setup. Use it only for authorized research and with the target's consent where required. 623
Verdict: star it if you study OSINT tooling or need a testable module pattern. Skip it as a production identity signal; service-side changes will outrun your assumptions.
6. cactus-compute/needle
Needle 2 attacks the hardware floor directly. It is a 45-million-parameter tool-calling and structured-extraction model packaged as a single 14 MB binary, with a full session designed to run in about 28 MB of RAM. Its target is a phone, wearable, smart-home device, or robot that needs structured local actions but cannot host a conventional language model. 24
The package combines a small-model recipe with a strict tool contract. The README describes a Simple Attention Network, GQA, an engram key-value memory, multi-lane hyper-connections, two-bit quantization, byte-level grammars compiled from schemas, a 256-token sliding window, and retrieval that narrows a large tool catalog to five candidates per turn. Python bindings cover inference, structured extraction, export, and LoRA fine-tuning; tuned models remain single
.cact files. 24Compared with a compressed general model, Needle makes tool use and bounded memory first-class constraints. That is its reusable pattern—and its limit. A 256-token window, a small tool subset, and a calibrated confidence threshold suit narrow actions, not open-ended conversation. The API shows 27 open issues, 13 contributor accounts, a last push on Aug. 15, and MIT licensing. The README includes API, playground, fine-tuning, data-format, and offline-device documentation; testing is not separately described. 725
Verdict: star it if you ship edge devices or want to study constrained agent inference. Skip it if you need broad conversational reasoning.
7. macro-inc/macro
Macro tries to remove the seams between the systems a small team uses every day. Email, messages, docs, tasks, agents, calls, files, pull requests, and CRM objects share one searchable workspace and one permission model. Its central claim is structural: cross-references are stored as a bidirectional graph, so a conversation can link to the task, customer record, or pull request that follows it. 26
The implementation uses SolidJS and Rust, with Markdown-native collaborative docs built on CRDTs and Cloudflare Durable Objects. The workspace exposes MCP and agent actions, while @-links connect surfaces that normally live in separate products. The result is closer to a team operating system than to another chat client: an email can create a task, a task can retain its originating context, and an agent can search across the same objects. 26
Macro differentiates itself from the Slack-plus-Linear-plus-Notion stack by making shared state the product instead of synchronizing several products after the fact. The cost is a very large promise and a new system to trust with company data. The API shows 68 open issues, 23 contributor accounts, a last push on Aug. 15, and AGPL-3.0 licensing. The README documents the major surfaces and examples but does not list a separate CI/test setup. 827
Verdict: star it if unified, agent-readable team state is the problem. Skip it if your team needs a narrow, mature point tool rather than a new operating layer.
8. vitali87/code-graph-rag
Code-Graph-RAG replaces file-by-file retrieval with a graph of a repository's structure. It parses functions, classes, methods, modules, and relationships across a mixed-language codebase, then lets an agent query, edit, detect dead code, and perform structural rewrites. The latest README also describes runtime call tracing, so test or production execution can add edges that static analysis missed. 28
The stack is Tree-sitter for parsing, Memgraph for the knowledge graph, Qdrant for the packaged semantic-search path, and a Python CLI that turns natural-language requests into Cypher queries. It supports Python, TypeScript, JavaScript, Rust, Go, Java, C/C++, C#, PHP, Lua, and Dart, with Ruby available through an ast-grep tier. An MCP server exposes the graph to Claude Code and other clients, while AST-based patches provide a diff before editing. 28
Its advantage over text RAG is relational context: callers, imports, symbols, and runtime edges can be queried together. The trade-off is infrastructure. Docker, Memgraph, CMake, and ripgrep are prerequisites, and a shared graph means destructive operations need care. The API shows 49 open issues, 49 contributor accounts, a last push on Aug. 16, and MIT licensing. The README has extensive installation, quickstart, architecture, language, MCP, and troubleshooting docs; testing appears as an input to tracing, not as a separate CI inventory. 929
Verdict: star it if agents need to understand a monorepo before editing it. Skip it when a small codebase makes a graph database heavier than the problem.
9. ToolJet/ToolJet
ToolJet is the chart's least specialized project and one of its most practical. It provides a visual builder for internal tools, workflows, and AI agents, with 60-plus UI components, a built-in database, JavaScript and Python execution, and more than 80 data-source integrations. The community edition can be self-hosted; the README separates it from enterprise AI features such as app generation, query building, debugging, and agent orchestration. 30
The implementation choice is a broad deployment surface rather than a single framework trick. Docker, Kubernetes, Helm, AWS, GCP, Azure, and other paths are documented, while plugins and connectors extend the platform. Tutorials cover a time tracker, a CMS, and an S3 browser. That makes ToolJet different from a code-generation demo: the output is a deployed application connected to real data and permissions, with an optional AI layer on top. 30
The trade-off is governance and maintenance at platform scale. The API shows 1,131 open issues, at least 409 contributor accounts across five returned contributor pages, a last push on Aug. 14, and AGPL-3.0 licensing. The README has deployment, self-hosting, examples, and contribution documentation, but its CI/CD references describe a product capability rather than a complete project-test inventory. 103132333435
Verdict: star it if you need self-hosted internal apps and can accept a platform-sized dependency. Skip it if your target is a tiny single-purpose tool.
10. addyosmani/agent-skills
agent-skills packages senior-engineering workflows for coding agents instead of shipping another coding agent. Its 24 skills cover a lifecycle from /spec and /plan through /build, /test, /review, and /ship, plus web performance, simplification, and context engineering. The skill files are plain Markdown and can be installed through the skills CLI or native integrations for many agent clients. 36That makes it a distribution layer. The reusable unit is a workflow contract with quality gates, not a model, runtime, or hosted service. Users can install the full pack or select a single skill; the README calls out a portability caveat when a per-skill install omits shared references. That caveat matters more than the number of supported clients: package boundaries determine whether a skill still works after it leaves the source repository. 36
The API snapshot shows 106 open issues, 62 contributor accounts, a last push on Aug. 14, and MIT licensing. The README has installation guides for multiple clients, an adoption guide, and a complete skill inventory. It describes testing as a workflow principle and quality gate, but does not provide a separate repository-level CI/test summary in the fetched page. 1137
Verdict: star it if you want repeatable engineering habits inside several agents. Skip it if you are looking for a runnable autonomous system; this repository supplies the operating instructions, not the runtime.
What this week says about open-source tooling
The chart's strongest projects make a messy capability addressable:
- Make the boundary explicit.
diagram-designconstrains visual output; Switchyard constrains provider routing; Needle constrains tool calls and memory;agent-skillsconstrains engineering behavior. - Store relationships, not just text. Semantica records provenance and decisions, Macro links team objects, and Code-Graph-RAG links code structure and runtime behavior.
- Treat the failure boundary as part of the product. Prime Agent is not a sandbox, holehe depends on changing service behavior, ToolJet carries platform-scale operations, and Switchyard is pre-alpha.
The stars worth carrying forward depend on the problem you actually have. For agent infrastructure, Prime Agent, Semantica, Switchyard, Code-Graph-RAG, and
agent-skills expose different layers that are easy to confuse: runtime, governed context, routing, code understanding, and workflow distribution. For constrained devices, Needle is the clear specialist. For teams, Macro and ToolJet make opposite bets on how much of the workspace to own. A GitHub star is useful here as a bookmark for a design pattern; it is not a production approval.References
- 1GitHub Trending weekly page
github.com
- 2diagram-design API metadata
api.github.com
- 3semantica API metadata
api.github.com
- 4prime-agent API metadata
api.github.com
- 5Switchyard API metadata
api.github.com
- 6holehe API metadata
api.github.com
- 7Needle API metadata
api.github.com
- 8Macro API metadata
api.github.com
- 9Code-Graph-RAG API metadata
api.github.com
- 10ToolJet API metadata
api.github.com
- 11agent-skills API metadata
api.github.com
- 12diagram-design repository
github.com
- 13diagram-design contributors
api.github.com
- 14Semantica repository
github.com
- 15semantica contributors
api.github.com
- 16Prime Agent repository
github.com
- 17prime-agent contributors page 1
api.github.com
- 18prime-agent contributors page 2
api.github.com
- 19prime-agent contributors page 3
api.github.com
- 20Switchyard repository
github.com
- 21Switchyard contributors
api.github.com
- 22holehe repository
github.com
- 23holehe contributors
api.github.com
- 24Needle repository
github.com
- 25Needle contributors
api.github.com
- 26Macro repository
github.com
- 27Macro contributors
api.github.com
- 28Code-Graph-RAG repository
github.com
- 29Code-Graph-RAG contributors
api.github.com
- 30ToolJet repository
github.com
- 31ToolJet contributors page 1
api.github.com
- 32ToolJet contributors page 2
api.github.com
- 33ToolJet contributors page 3
api.github.com
- 34ToolJet contributors page 4
api.github.com
- 35ToolJet contributors page 5
api.github.com
- 36agent-skills repository
github.com
- 37agent-skills contributors
api.github.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.