#agents

Curated list of open-source AI frameworks and platforms for infrastructure, fine-tuning, research agents, conversational AI, education, computer vision, edge computing, and finance.

Curated list of sandboxes, runtimes, orchestration platforms, and infrastructure frameworks for running AI agents.

Agent Reach (Panniantong/Agent-Reach, MIT) is a Python CLI that gives coding agents read commands for web and social channels. Checked 3 November 2026, version 1.6.1 matches the latest release, and the PyPI package agent-reach is a different project.

Overview of multi-agent tools and frameworks, covering Make It Heavy, Awesome LLM Apps, CrewAI, and a list of popular AI agent frameworks.

Curated list of agent skills, context-engineering and compression tools, browser automation for agents, and other productivity tools for coding agents.

anymd reads a public URL into structured Markdown, stores it in a private library, and serves it through a URL prefix, REST, CLI, and a remote MCP server.

How the accessibility tree is built, how to keep a UI's tree small and named, and how an agent should prune it before the model sees it.

BrowserOS (browseros-ai/BrowserOS, AGPL-3.0) is a Chromium monorepo that ships BrowserOS neo for agents and BrowserOS for people. Checked 3 November 2026, the newest GitHub release is BrowserOS neo v0.52.0, and neither app was installed.

Cloudflare's open-beta Web Search API sends a query through AI Gateway to Ceramic.ai, Exa, or Linkup and returns one result shape, billed at each provider's list price.

Curated overview of AI coding agents, CLIs, VS Code extensions, IDEs, and app builders such as OpenCode, Devin, Cosine, Dyad, and Kilo.

eve (vercel/eve, npm eve 0.80.1, Apache-2.0, public beta) is Vercel's filesystem-first TypeScript framework for durable backend AI agents - an agent is an agent/ directory of instructions, tools, skills, connections, channels, schedules and subagents; sessions run as Workflow SDK durable workflows with a per-session sandbox, evals, useEveAgent frontend hooks, and Agent Runs tracing on Vercel or self-hosted on Node.

Google's pre-stable agent orchestration runtime. kubectl-style Task, Workspace, and Model manifests, sandboxed as actors on Agent Substrate, with checkpoint suspend and resume.

gstack (garrytan/gstack, MIT) is Garry Tan's skill pack that adds review, QA, and release commands to coding agents. Checked 3 November 2026, main VERSION is 1.93.21.0, the newest GitHub release is v1.92.55.0, and the pack was not installed.

ktx (npm @kaelio/ktx 0.16.0, Kaelio/ktx, Apache-2.0) is a local context layer that turns a SQL warehouse and team docs into a wiki and semantic layer, then serves approved metrics to coding agents over a CLI and MCP; the unscoped npm package ktx and the Khronos KTX texture format are different projects.

Curated list of frameworks and platforms for building multi-agent systems, including Agno, TinyTroupe, Google ADK, Dify, and TradingAgents.

Nuphos (nuphos/nuphos, Apache-2.0, by Zeabur Inc.) is a team workspace for coding agents - Claude Code, Codex, Grok Build, Antigravity and OpenCode run over ACP on local, remote or cloud runtimes, with shared multi-user sessions, team skills, memory and credentials, DevOps GUIs (Kubernetes, Grafana, cloud consoles), Slack/Discord, triggers, and Electron/iOS/CLI clients. Early access; the Desktop app ships PostHog analytics with session replay on by default.

OpenMausBot (milind-soni/OpenMausBot, Apache-2.0) is a local-first chat app that runs a team of coding-agent bots. Checked 3 November 2026, main is package 0.1.139, the newest GitHub release and npm latest are 0.1.138, and the app was not installed.

OpenWorker (andrewyng/openworker, MIT) is a local desktop AI coworker with a Python server and a Tauri shell. Checked 3 November 2026, the Tauri config and Python package on main say 0.3.4, the newest GitHub release is v0.3.1, and PyPI openworker is 0.0.1.

Experimental durable agent harness from Earendil (shipped with Pi 1.0) for long-running, crash-survivable, multi-conversation agent apps over memory/JSONL/SQLite storage, with pluggable extensions, tools, hooks, tasks, compaction, and documents.

REA (npm rea-agents 6.8.0, MIT) is a local CLI and MCP server that returns evidence for native binaries, JavaScript, Electron, .NET, Android, firmware, browsers, and other static formats. Checked 2 November 2026: GitHub lists 420,599 stars, the site is rea.tools, and main has commits after the 25 October 2026 npm release.

Skill Recorder (microsoft/skill-recorder, MIT, Electron + TypeScript) records a screen session on-device, optionally transcribes narration with Whisper and blurs OCR-located text before anything leaves the machine, has GitHub Copilot reconstruct the intent and ordered steps, then renders a SKILL.md into ~/.copilot/skills or a Scout automation. 5.4k stars, v0.7.0.

treg is an open-source tool catalog and credential-injecting proxy for AI agents — one token for 4,000+ priced API endpoints across 121 providers, plus a team's own keys, CLIs and skills, with $1 free credit and 0% markup.

Open-source AI editor copilot plugin for Unreal Engine 5.6 that runs as a dockable in-editor chat tab, drives the editor via Python execution, scene queries, viewport screenshots, and optional MCP/Replicate tools.

Native local-first control plane (Rust, gpui) for driving coding agents such as Claude Code, Codex, Cursor, Devin, Grok, Hermes, Pi, and Antigravity, with optional multi-device sync over Loro and Cloudflare Durable Objects.

Laravel's first-party AI SDK (laravel/ai) — a unified PHP API for agents with tools and structured output, images, audio, transcription, embeddings, reranking, classification, vector stores, and human tool approval.

NovelKit (novelkit.cc) is a Vietnamese long-form fiction studio; the public V2 Lite repo is a local single-operator pipeline under a source-available non-commercial, no-derivatives license, and it is not NovelAI.

Notes on an adaptive A/B testing tool architecture that uses Amazon Bedrock and MCP to assign variants based on user context.

Learn Harness Engineering is WalkingLab's MIT course on the files, checks, and session rules around a coding agent. The English site has 14 lectures and 8 projects, and the repo was last pushed on 22 October 2026.

Cloudflare's next-generation agentic CLI (cf) covering the entire Cloudflare API via OpenAPI-generated commands, cloudflare.config.ts TypeScript config, and Vite-first local development — the intended successor to Wrangler.

Hallmark (Nutlope/hallmark, MIT, v2.0.0, made by Together AI) is a design skill for Claude Code, Codex, Cursor and OpenCode that builds new pages from one of 48 dealt themes under a non-negotiable Floor of design rules, then lints the result in a real browser with sloplint --render and a fresh-context review. 35.7k stars, 1.9k forks, 62 open issues at check time.

ptn1411/skill ("Authorized Artifact Auditor", Python, no license file) is a pack of about 35 agent skills plus Python CLIs for reverse engineering and auditing software you own - fingerprinting, unpacking, decompiling, web and network recon, SBOM, CVE lookup. The README and MASTER_POLICY.md say defensive use only, but several skills in the repo (master-unlock, dotnet-keygen, binary-patcher, frida-hooker) are written as bypass and keygen tools. Read before installing.

alexgaoth/awesome-reverse-engineered is a searchable index of about 12,900 open-source projects that exist because someone reverse engineered a closed app, API, device, protocol, file format, game or AI tool, with a website, a JSON dataset (CC0) and an agent skill that searches it. Use it to check whether the work is already done before reverse engineering something yourself.

RedAmon (samugit83/redamon, MIT) is a self-hosted Docker penetration-testing framework that maps an authorized target into a Neo4j graph, runs an agent from a Kali sandbox behind approval gates, and can open a GitHub pull request for a fix.

security-audit (cloudflare/security-audit-skill, MIT) is an agent skill that runs a six-phase defensive security audit - reconnaissance with a coverage ledger, isolated hunters, adversarial candidate validation, schema-validated findings.json, independent record verification, and target-neutral reports - seeded from Cloudflare's fleet-wide vulnerability discovery harness. 48.2k stars, 3k forks, 98 open issues at check time.

Notes on Alfonso Campodonico's "Building your own sandboxes" (2026-10-18) - vmkit runs Firecracker or Cloud Hypervisor microVMs behind one API, each VMM inside a rootless namespace jail with a default-deny per-VM network, and kiln turns OCI images into one-erofs-per-layer microVM images. Main lesson - the VMM is the easy part, the policy around it is the work.

e2e (npm e2e 0.18.0, tester-army/e2e, Apache-2.0) is TesterArmy's pre-1.0 web and mobile runner that mixes natural-language agent steps with locator assertions, using Playwright in the browser and agent-device on devices; npm versions 0.0.x are the older willscott/e2e OpenPGP library.