24 Open Source Projects Worth a Look Right Now

Memory, security, and token savings, all in one list

Photo by Towfiqu barbhuiya on Unsplash

The AI tooling space is throwing off new releases faster than anyone can read the changelogs. Some of what follows quietly fixes the token bill nobody wants to talk about. Some of it stops an agent from re-explaining a codebase it already knows. One tool exists purely to catch malicious code hiding inside a skill package before it does damage. Another turns a pile of messy web pages into something an LLM can actually use without choking on the HTML.

Here’s a rundown of 24 projects spanning orchestration, memory, security, and infrastructure. Grab the link, star what’s useful, skip the rest.

Haystack

Haystack is an open-source framework for assembling LLM-powered applications out of modular parts: document stores, embedders, retrievers, and agent pipelines. Rather than building a RAG system from raw API calls, developers snap together pre-built components and swap them as requirements change. deepset has maintained the project for years, so it ships with mature documentation and a real production track record.

• What you get: a production-tested foundation for RAG and agent pipelines without reinventing the retrieval stack.

Recall

Recall gives Claude Code durable memory of a project so context doesn’t disappear the moment a session ends. Instead of re-pasting file structures or past decisions at the start of every chat, the tool holds that knowledge and feeds it back automatically. It’s built for anyone running long Claude Code sessions who is tired of paying token costs just to restate the obvious.

• What you get: an end to re-explaining your own codebase every time you open a new session.

Knowledge Work Plugins

This is Anthropic’s own open-source collection of plugins built for knowledge workers using Claude Cowork. Each plugin packages a specific task pattern, research, drafting, or review, into something Cowork can run directly instead of assembling from scratch. Because Anthropic maintains the repo, it also doubles as a reference for how first-party plugins are meant to be structured.

• What you get: a ready-made plugin library plus a blueprint for building your own.

Pulumi do

Pulumi do adds a direct command for creating, reading, updating, and deleting individual cloud resources without writing a full infrastructure program first. It’s aimed at the moments a proper IaC project doesn’t fit: quick fixes, one-off resources, or debugging drift where spinning up an entire stack is overkill. The command spans every cloud Pulumi already supports, so the shortcut doesn’t cost any coverage.

• What you get: a way to touch one cloud resource without the ceremony of a full IaC program.

mvm

mvm is an interpreter and virtual machine built for fast execution of Go and other languages. Speed is the entire design brief here, not a general-purpose toolchain replacement. For anyone experimenting with language runtimes or embedding a scripting layer in an app, it’s a lighter option than standing up a full compiler.

• What you get: a fast, embeddable runtime for projects that need scripting without full compilation overhead.

codegraph

codegraph builds a pre-indexed knowledge graph of a codebase and keeps it synced automatically as the code changes. It plugs into Claude Code, Codex, Gemini, Cursor, and OpenCode, letting several AI coding tools share the same map of how a project fits together. That shared index cuts down on re-scanning files and improves the accuracy of answers about how pieces of the code connect.

• What you get: one synced code graph that every AI coding assistant in your stack can query.

herdr

herdr is a terminal-based multiplexer for running several coding agents side by side. It lets a developer track and control parallel agent sessions from a single terminal window instead of switching between tabs or windows. The project suits anyone running multiple agents at once who wants one control point instead of several.

• What you get: a single terminal view for managing multiple coding agents at once.

OmniRoute

OmniRoute is a free AI gateway that routes requests to more than 231 providers, over 50 of them free, through a single endpoint. It connects tools like Claude Code, Codex, Cursor, and Cline so a developer can switch providers without rewriting integration code. The pitch is simple: a coding agent never runs dry because one endpoint always has a fallback.

• What you get: an endpoint that keeps coding agents running even when a favorite provider is capped or down.

no-mistakes

no-mistakes is a lightweight git tool built around the idea of catching problems before a push goes out. Its scope is narrow by design: prevention over cleanup. Details in the repo are sparse, which makes this one read as an early-stage project worth watching rather than a finished tool.

• What you get: a minimal safeguard against pushing mistakes to a shared branch.

Orca

Orca is an agent development environment for running a fleet of parallel coding agents. It works with any coding agent under a developer’s own subscription instead of locking users into one vendor. The goal is coordinating many agents at once rather than babysitting one agent through one task at a time.

• What you get: the ability to run a whole fleet of coding agents in parallel instead of one at a time.

agentsview

agentsview is a local-first tool for searching sessions, tracking analytics, and measuring token use across coding agents. It supports Claude Code and similar tools, giving a developer real visibility into what an agent did across past sessions. Because it runs locally, the usage data stays on the developer’s machine instead of a hosted dashboard.

• What you get: local session search and token analytics without sending usage data anywhere else.

SkillSpector

SkillSpector is a security scanner built specifically for AI agent skills. It checks for vulnerabilities, malicious patterns, and other security risks that could be hiding inside a skill package before it’s ever installed. Coming from NVIDIA, it signals growing attention to supply-chain security as agent ecosystems keep expanding.

• What you get: a scan for malicious code in agent skills before they get anywhere near production.

aisuite

aisuite is a simple, unified interface for calling multiple generative AI providers through one consistent API shape. Andrew Ng’s team built it to remove the friction of learning a new SDK every time a project adds another provider. It trades some provider-specific features for consistency, which suits projects that need to swap models without rewriting call sites.

• What you get: one consistent API for calling many different AI providers.

Meshery

Meshery is a cloud native management platform focused on service meshes and the infrastructure around them. It gives teams a single place to configure, visualize, and manage cloud native tooling instead of juggling separate dashboards per tool. The project has years of development and an active community behind its current feature set.

• What you get: a unified control plane for managing cloud native infrastructure and service meshes.

opencode

opencode is an open-source coding agent built as an alternative to closed, proprietary agents. It’s designed to be extended and modified freely, which appeals to developers who want to inspect or change how their coding agent actually behaves. Being open source also means it can run without any ties to a specific vendor’s roadmap.

• What you get: a fully open coding agent you can inspect, modify, and self-host.

markitdown

markitdown is a Python tool that converts files and Office documents into clean Markdown. It handles common formats like Word, PowerPoint, and PDF, turning them into something both humans and LLMs can read easily. Microsoft built it with LLM pipelines in mind, so the output stays structured rather than cluttered with formatting artifacts.

• What you get: clean Markdown out of messy Office files, ready to feed straight into an LLM.

Crawl4AI

Crawl4AI is an open-source web crawler and scraper built to produce output that’s friendly to LLMs. It scales from a single page fetch to a full site crawl, adjusting its approach based on how much ground it needs to cover. An active Discord community backs the project, which matters for a tool that runs into as many edge cases as web scraping does.

• What you get: a scraper that hands back LLM-ready content instead of raw, messy HTML.

zvec

zvec is a lightweight, in-process vector database built for speed. It runs embedded inside an application rather than as a separate service, cutting out network latency for vector search. Alibaba built it for the cases where a full-scale vector database is more infrastructure than the problem actually calls for.

• What you get: vector search that lives inside your app instead of requiring a separate database service.

codebase-memory-mcp

codebase-memory-mcp is an MCP server that indexes a codebase into a persistent knowledge graph. It’s built for speed, with claims of indexing an average repo quickly, and for staying current as the code keeps changing. The result is code intelligence that any MCP-compatible AI tool can query directly instead of re-reading files from scratch.

• What you get: fast, persistent code intelligence exposed as an MCP server any compatible tool can query.

pi-subagents

pi-subagents is a Pi extension for delegating tasks to async subagents. It supports truncation, artifacts, and session sharing, so a subagent can hand back a manageable result instead of flooding the parent session with raw output. The project targets developers who want to offload work without losing track of what those subagents actually produced.

• What you get: delegated work that comes back as clean, shareable results instead of a wall of raw output.

Supermemory

Supermemory is a memory and context engine, paired with a companion app, built for speed and scale. It can run fully locally, which matters for anyone who doesn’t want their memory layer dependent on a hosted API. The project positions itself as the memory API for the AI era, built for applications that need to remember things across sessions.

• What you get: a memory layer for AI apps that can run entirely on your own infrastructure.

headroom

headroom compresses tool outputs, logs, files, and RAG chunks before they ever reach an LLM. The project claims 60 to 95 percent fewer tokens while keeping answer quality steady. That kind of compression matters most for anyone running high-volume agent setups where token costs add up fast.

• What you get: a 60 to 95 percent cut in token usage before that content ever hits the model.

Scrapling

Scrapling is an adaptive web scraping framework built to handle everything from a single request to a full-scale crawl. It’s designed to adjust automatically as target sites change, reducing the maintenance burden that usually comes with scraping. That adaptability is the main draw for anyone who’s had a scraper break the moment a site redesigned its markup.

• What you get: a scraper built to survive site redesigns instead of breaking every time the markup shifts.

turbovec

turbovec is a vector index built on TurboQuant, written in Rust with Python bindings. Rust gives it a performance edge, while the Python bindings keep it accessible to the ecosystem most AI development already lives in. It’s a narrower, more specialized tool than a full vector database, built for developers chasing raw indexing speed.

• What you get: Rust-speed vector indexing with a Python interface most AI developers already use.

Bottom Line

A few threads run through this list even though the projects don’t talk to each other. Some exist to stop AI tools from forgetting what they already learned about a project. Some exist to keep a fleet of coding agents running in parallel without one person watching every terminal. One quietly shrinks the token bill before a single extra dollar gets spent. Another builds a shared map of a codebase so every AI tool touching that code sees the same thing. And one just checks that the skill you’re about to install isn’t secretly malicious.

None of these require a platform migration or a rewrite to try. Pick the one solving today’s actual problem, star it, and move on.