27 things to install into your agent this week: a quarter-million-star self-improving loop, a kernel-level sandbox built by the Sigstore team, and a run of skill packs and MCP servers that only showed up in the past month.
Over 1,500 skills sit in this directory now, curated from real engineering teams at Anthropic, Microsoft, Google, and Vercel rather than mass-generated filler, so it's worth checking before you write your own from scratch. Each entry links out through officialskills.sh to that skill's actual install steps, which vary by author. Pricing: free, open source.
npx skills add vercel-labs/agent-skills drops in Vercel's own pack: forty-plus React and Next.js performance rules, a hundred-plus accessibility checks pulled from their web-design guidelines, and a writing-guidelines skill that enforces their own prose standards. Pricing: free, open source.
R and Python package developers get their own pack here: Shiny app theming, Quarto publishing, R package lifecycle management, and a ggsql skill for writing SQL through plain language, added September 14, 2026. npx skills add posit-dev/skills --all installs the full set, or /plugin marketplace add posit-dev/skills to pick categories individually. Pricing: free, open source.
Anthropic runs its own plugin directory now, split into an internal /plugins folder it maintains directly and an /external_plugins folder for vetted partner and community work, each one bundling commands, agents, skills, and MCP config together. /plugin install <name>@claude-plugins-official is the whole install. Pricing: free, open source.
buildwithclaude.com indexes 175 commands and 117 agents directly, then points past its own catalog into more than 20,000 community plugins and 4,500 MCP servers it doesn't host itself, working more like a search engine for the rest of this list than a library of its own. /plugin marketplace add davepoon/buildwithclaude adds it, then install whatever turns up. Pricing: free, open source.
PRP stands for Product Requirement Prompt, and the dozen commands here (/prp-plan, /prp-implement, /prp-review, /prp-orchestrate) carry a feature from spec through validation instead of leaving that structure up to you. A PR-triage command landed September 8, 2026 alongside a shorter repo name. Install with /plugin marketplace add Wirasm/PRPs-agentic-eng then /plugin install prp-core@prp-marketplace. Pricing: free, open source.
Google's Chrome team maintains this one directly: real browser control through Puppeteer, DevTools performance traces, network inspection, and console output an agent can read without a screenshot. One mcp.json entry pointing at npx -y chrome-devtools-mcp@latest is the entire setup, and a --slim --headless flag trims it for simple tasks. Pricing: free, open source.
Cloudflare rebuilt every server in this repo on a new stateless SDK factory this year, so each one answers through the same /mcp endpoint instead of holding session state in memory, and the old SSE URLs now return a 410 with migration instructions. Point an MCP client at a remote URL like https://docs.mcp.cloudflare.com/mcp for documentation or https://bindings.mcp.cloudflare.com/mcp for Workers bindings; nothing runs locally. Pricing: free, usage-based on the Cloudflare resources it touches.
Everything else in this bucket is a single server. Agentgateway is the layer that sits in front of all of them, a Rust proxy that adds OAuth, content filtering, and observability to MCP and A2A traffic without touching the servers themselves. Standalone and Kubernetes quickstarts are both documented at agentgateway.dev. Pricing: free, open source.
Augment indexes your whole codebase into a context engine before its Agent touches anything, aimed at repos too large for a model to hold in one context window. It's a normal VS Code Marketplace install, already past 778,000 of them. Pricing: free extension, requires a paid Augment plan for agent usage.
Nanobrowser puts a small team of agents inside a Chrome extension instead of a server you run separately, splitting a web task like scanning GitHub Trending or comparing Amazon listings across specialized roles, and it works with whatever LLM key you already hold: OpenAI, Anthropic, Gemini, or a local Ollama model. Install from the Chrome Web Store, or load the unpacked build from GitHub releases for the newest features. Pricing: free, open source, bring your own API key.
Companion for Claude turns an Obsidian vault into Claude's working directory: vault-aware chat, an agent mode that can call tools, and planning artifacts that sync straight back into your notes once you approve the edit. It shipped version 0.29.1 within the past week and is still under 2,000 downloads, early enough that feedback is still shaping it. Search "Companion for Claude" in Obsidian's Community Plugins browser to install. Pricing: free, open source, bring your own Claude API key.
Most loops in this bucket repeat a fixed plan; Hermes writes and refines its own skills as it works, then nudges itself to save what it learned so the next session doesn't start from zero. It launched February 25, 2026 and has since passed a quarter-million stars, the fastest growth of anything in this space this year. curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash installs it, and it'll connect to Telegram, Discord, Slack, WhatsApp, or Signal if you want it reachable outside a terminal. Pricing: free, open source; model API costs are separate.
Agent Canvas is OpenHands' control-center product: a self-hosted dashboard for running several coding agents as an always-on team, wired into Slack, GitHub, and Linear so a task can start from a ticket instead of a terminal command. npm install -g @openhands/agent-canvas followed by agent-canvas launches the full local stack. Pricing: free, open source.
Omnigent doesn't run any agent itself. It sits above Claude Code, Codex, Cursor, OpenCode, Hermes, and Pi as a shared policy and sandboxing layer, syncing a session across devices in real time. Version 0.14.0 shipped September 15, 2026, though the project still labels itself alpha. Install with curl -fsSL https://raw.githubusercontent.com/omnigent-ai/omnigent/main/scripts/install_oss.sh | sh. Pricing: free, open source.
OpenCode ships two built-in agents in one CLI: a build agent with full write access and a plan agent restricted to read-only analysis, backed by 75-plus model providers and LSP integration so it reads a project the way an editor would. It moved from the sst GitHub org to anomalyco this year, but the install command hasn't changed: curl -fsSL https://opencode.ai/install | bash. Pricing: free, open source; model API costs are separate.
ccmanager skips tmux entirely and builds its own TUI for running Claude Code, Gemini CLI, Codex CLI, Cursor Agent, Copilot CLI, Cline, OpenCode, and Kimi CLI side by side in separate git worktrees, with a live busy, waiting, or idle indicator per session so you know which one actually needs you. npm install -g ccmanager, or skip the install and run npx ccmanager directly. Pricing: free, open source.
Conductor is a Mac app rather than a CLI wrapper: it fires Claude Code, Codex, and Cursor at the same task in separate git worktrees, then lets you diff and merge or discard each result, with a Linear ticket as an optional starting point. Its changelog logged a version bump the same day this digest was written. Download it from conductor.build. Pricing: free.
Guardrails AI wraps a validator function around an LLM call to catch what shouldn't reach a user or a downstream system: PII, jailbreak attempts, toxic language, hallucinated claims, pulled from a hub of pre-built validators instead of a regex you write yourself. pip install guardrails-ai installs the framework; validators come from the Hub separately. Pricing: free, open source; hosted inference was retired August 25, 2026, so it now runs fully self-hosted.
Microsoft built this to make a policy violation structurally impossible rather than trust a model to refuse on its own: it intercepts a tool call before execution and checks it against policy, with claimed coverage of all ten OWASP Agentic AI Top 10 risks plus NIST and EU AI Act mappings. pip install "agent-governance-toolkit[full]" covers Python; TypeScript, .NET, Go, and Rust packages ship separately. Pricing: free, open source.
Nono comes from the team behind Sigstore and skips containers entirely, using Linux Landlock or macOS Seatbelt to strip an agent's process privileges for the run, then sandboxing each tool it calls, git, curl, kubectl, in its own separate child sandbox so a compromised agent can't inherit more access than its policy allows. Trade coverage picked it up as an open-source AI agent sandbox on July 27, 2026. Install with curl -fsSL https://nono.sh/install.sh | sh. Pricing: free, open source.
Mem0 keeps memory at three levels, user, session, and agent, and retrieves by similarity instead of replaying a full transcript, which is why it's become the default answer whenever a tutorial needs "an agent that remembers." pip install mem0ai or npm install mem0ai gets the library; a hosted version and CLI exist too. Pricing: free, open source; managed hosting is a paid tier.
claude-mem watches every tool call in a Claude Code session, compresses what happened with a smaller model, and quietly reloads the relevant parts the next time you start working, so a session doesn't begin from nothing. It grew past 94,000 stars by the middle of this year and now reaches past Claude Code into OpenCode, Grok Bot, and Antigravity. npx claude-mem install sets it up, with an --ide flag for the others. Pricing: free, open source.
GoodMemory is the small, legible option next to Mem0: memories live in plain SQLite, every one can be inspected, corrected, or deleted through a forget call, and nothing writes back to storage unless you opt in. It shipped its first stable release July 6, 2026 and reached v0.8.0 on September 14 with MCP tool exposure added. npm install -g goodmemory@0.8.0 then goodmemory setup wires it into Codex or Claude Code. Pricing: free, open source.
Langfuse covers the whole loop an eval setup needs: tracing every LLM call, versioning prompts, running LLM-as-judge or custom evaluators against a dataset, and a playground for iterating without redeploying anything. git clone plus docker compose up gets a self-hosted instance running locally, with Kubernetes and cloud templates for production. Pricing: free, open source core; managed cloud and enterprise features are paid.
Opik chases the same job as Langfuse but leans harder into agent-specific tracing, following multi-step tool calls through sixty-plus framework integrations, and it stayed fully self-hostable under Apache 2.0 even after a 2026 Gartner report named it in its market guide for AI evaluation platforms. pip install opik && opik configure gets the SDK; ./opik.sh after cloning runs the full self-hosted stack. Pricing: free, open source; Comet's managed cloud is a paid tier.
Everything upstream in this bucket traces LLM calls; hwatu checks whether the resulting UI actually rendered right. A WebKitGTK daemon opens a window in about 13 milliseconds, screenshots it, and scores the pixel difference against a baseline so an agent can iterate on a fix instead of guessing, with a handoff to a human mid-session when it hits a CAPTCHA. It surfaced on Hacker News around July 29, 2026. Install with curl -fsSL https://raw.githubusercontent.com/hongnoul/hwatu/main/scripts/install.sh | bash. Pricing: free, open source.