A diagramming skill for Claude Code, Codex, and Pi that outputs self-contained HTML + SVG, with dedicated templates for architecture sketches, flowcharts, and pyramids.
Stars 19,271This week +14,735Forks 1,170Open issues 22Last commit 2026-08-14
It encodes what makes a diagram good as rules, so model output is finally publishable as-is. It is a natural fit if you already use these coding agents to write technical posts, product docs, or internal proposals. The catch is that it produces static HTML/SVG, so it does not slot neatly into a workflow that finishes every diagram in Figma.
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. Output is self-contained HTML + SVG with no build step and no external image dependency, so it drops straight into a site (the README says 27 visual types, the repo description says 29 — check the actual directory). Sitting on old draw.io or Mermaid sources? It'll redraw them at a size and level of detail you pick.
Gaining 14,700 stars in a week says the thing it hit isn't diagrams — it's the floor on how agent output looks. The same crowd has had enough of shipping work that reads as machine-made.
Prime Intellect's open-source coding and research agent, built on a recursive language model and a durable harness for long-running work.
Stars 16,472This week +8,488Forks 1,774Open issues 74Last commit 2026-08-16
Context as variables and subagents as function calls let effective working patterns survive past one session. It makes the most sense for refactors, migrations, and research that keep running for hours or days. If all you want is autocomplete, orchestrating everything in code is far more machinery than you need.
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 arXiv:2605.09998), 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. 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.
Graph-native context and knowledge infrastructure that turns enterprise data into a context graph and knowledge graph, with decision provenance on every conclusion.
Stars 8,107This week +5,339Forks 827Open issues 70Last commit 2026-08-16
Billing itself as "the open source Palantir for AI agents," it targets the enterprise need to justify an answer. Teams in finance, healthcare, or public services that must explain model decisions to compliance are the clearest audience. It does assume real knowledge-modeling and ontology skills; for an internal Q&A bot, this is overkill.
A team-level memory hub for agents that distills conversations, docs, and code into four reusable assets: Chat Memory, Skill, LLM-Wiki, and Code-Graph.
Stars 22,148This week +3,956Forks 2,031Open issues 603Last commit 2026-08-15
Pulling memory out into a shared service lets several people and tools draw on the same team knowledge. Teams juggling several agent tools will get the most from it when they want durable team knowledge rather than another pile of personal chat logs. That comes with real setup: Node 22.16+, three self-hosted services, and a LICENSE file worth reading because GitHub does not recognize it as a standard license.
Production-grade engineering skills for AI coding agents, with eight slash commands spanning spec, plan, build, test, review, and ship.
Stars 87,661This week +3,300Forks 9,392Open issues 104Last commit 2026-08-14
It turns senior-engineer process and quality gates into skills the agent follows at every phase. Teams using Claude Code, Codex, or Cursor can use it to add the process they missed after an agent skipped design and went straight to code. The process is deliberately heavy, though; six steps for a one-line bug fix feels like a lot.
A 45M-parameter on-device model for tool calling, device use, and structured extraction, shipped as a single 14MB binary by the project's own account.
Stars 6,366This week +2,488Forks 424Open issues 25Last commit 2026-08-15
It squeezes tool calling into watch-sized memory, making offline on-device agents a real option. Its real audience is engineers building on-device, wearable, robotics, or offline products that need predictable JSON output. It is not a general chat model, so do not expect prose or hard reasoning; the benchmark figures also come from the project team itself.
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, with weights on Hugging Face and the 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. The comparisons against FunctionGemma 270M and LFM2.5 230M are the team's own numbers, so re-run them before you ship.
Email, chat, docs, tasks, calls, CRM, and agents in one workspace, @-linked over a single bidirectional graph.
Stars 3,348This week +2,434Forks 332Open issues 65Last commit 2026-08-16
Shared team-level memory means people and agents read the same context instead of MCP-and-Zapier glue. Startups of a few to a few dozen people will see the benefit only if they are willing to replace most of their existing stack. AGPL-3.0 is a hard constraint for closed-source derivatives or SaaS, and migrating away from an entrenched toolchain will not be cheap.
Open-source orchestration for teams of AI agents — a Node.js server and React UI for assigning goals and tracking work and costs.
Stars 78,431This week +2,430Forks 14,373Open issues 5,093Last commit 2026-08-16
It positions itself as the company around your agents rather than a single assistant, with cost tracking built in. Teams already running several agents will find it easiest to adopt when they need one place to assign work and watch spend. More than 5,000 open issues suggest growth is outpacing consolidation, so test stability carefully before production.
A desktop app for running and training LLMs and diffusion models locally, covering Qwen, Kimi, Gemma, DeepSeek, FLUX, and more.
Stars 72,340This week +2,207Forks 6,518Open issues 1,234Last commit 2026-08-16
It went from a tuning library to an install-and-go desktop app, lowering the bar for local fine-tuning again. Developers and researchers with a local GPU can use the desktop app as a gentler way into running or fine-tuning open weights themselves. The desktop build is still beta, and once VRAM gets tight the practical model list shrinks much faster than the catalogue suggests.
Parses a multi-language codebase with Tree-sitter into a Memgraph knowledge graph, then lets you query, edit, and optimize it in plain language.
Stars 4,408This week +1,756Forks 597Open issues 51Last commit 2026-08-16
The change worth noticing this week is runtime tracing: calls that actually happen during tests now flow back into the graph, revealing relationships static analysis misses. It earns the deployment cost when a team is about to refactor, document, or de-clutter a large multi-language monorepo. You still have to run Memgraph yourself; on a small repository, editor search is simply faster.