Open-Source AI Repos & MCP Servers

Browserbase just turned Claude Code into a real web agent. The browserbase/skills repo went live in October 2025 and exploded onto GitHub Trending the week of May 4-5, 2026 with a wave of new agent skills designed specifically for the Claude Code Agent SDK. 3,188 stars and 212 forks as of today. This is the production-grade browser-automation SDK that's been missing from the Claude Code ecosystem. You install the marketplace, pick the skills you need, and Claude gains real browser capabilities — not a sandbox, not a screenshot trick, but a full Browserbase remote session that can log into your accounts, scrape JavaScript-rendered pages, run adversarial UI tests, and deploy serverless browser functions. The 11 skills shipped today browser — CLI-driven web automation against remote Browserbase sessions. browserbase-cli — the official bb CLI for managing platform workflows from Claude. functions — deploy serverless browser automation to Browserbase's cloud, callable from any agent. site-debugger — diagnose and auto-fix failing automations. The skill that pays for itself in saved hours. browser-trace — capture full DevTools protocol traces. Critical for debugging flaky tests. safe-browser — local agent SDK with domain allowlist enforcement, so the agent can't wander off-policy. bb-usage — display real-time usage stats and cost forecasts inside Claude. cookie-sync — sync cookies from your local Chrome into a Browserbase session. The missing piece for "log in to authenticated pages" workflows. fetch — retrieve HTML or JSON from static pages. search — structured web search results, not just a markdown dump. ui-test — AI-powered adversarial UI testing. Feed it a git diff and it runs your changes against the live UI in a real browser. How to install One command inside Claude Code: /plugin marketplace add browserbase/skills Then install the browse plugin and restart Claude Code. Each skill is a folder under skills/ in the repo and each can be enabled independently. Why this matters The headline skill is ui-test. It takes a git diff, spins up a real browser, navigates to the affected page, and runs an adversarial test pass against your UI changes. That used to require a dedicated test engineer plus a CI pipeline plus Playwright/Selenium glue. Now it's one install. For frontend teams shipping multiple changes a day, this single skill changes the economics of testing. cookie-sync is the second sleeper hit. Headless browser agents have always struggled with authenticated pages — you either build a fragile credential injector, or you bake cookies into the agent's state. cookie-sync grabs them directly from your local Chrome profile. Claude can now legitimately log into your dashboards, scrape your data, file your tickets. How this differs from other Claude skills This is operational infrastructure, not opinionated configuration. Compare it to Matt Pocock's engineering skills (curated prompts and conventions for senior engineers) and Andrej Karpathy's single-file CLAUDE.md (prescriptive style guide). Those tell Claude how to think. Browserbase Skills give Claude what to do with a browser. Pairs with The parent platform: Browserbase — the cloud-browser infrastructure these skills run on. If you're not already on Browserbase, you'll want an account before installing — most of the skills require a remote session. Status Repo created October 12, 2025. Latest commit May 13, 2026. Active development from the Browserbase team. JavaScript primary language. No explicit license file in the repo as of today — the skills are published under Browserbase's standard terms, but if you're building commercial automation on top, confirm with their team.

mattpocock/skills is Matt Pocock's personal .claude directory — the curated engineering skills he uses with Claude Code every day, lifted directly from his daily config and shipped as a repo.This is the counterpoint to Karpathy's CLAUDE.md moment. Where Karpathy gave the community one prescriptive file and let it ricochet across every greenfield project, Pocock ships a full directory — multiple skills, each scoped to a specific engineering task. TypeScript skills, Vitest skills, refactor skills, library-evaluation skills. It's what a Claude Code config looks like when a senior engineer has actually lived with the tool for months.Why it spikedTwo reasons, both real:1. The community wants worked examples, not theory. Anthropic's skill docs explain the mechanism. Pocock's repo shows what a working skill actually contains — the prompt shape, the file references, the trigger criteria. People clone, adapt, and run.2. TypeScript developers are the highest-leverage Claude Code audience right now. Pocock built his reputation on TypeScript pedagogy. His skills are TS-flavored — type narrowing, generic constraints, library refactors. That's the same crowd already using Claude Code aggressively, so the fit is unusually tight.Installgit clone https://github.com/mattpocock/skills .claude # or pick individual files and drop into your existing .claude/skills/Read each skill before activating — they're opinionated about TypeScript style and may not match every codebase.What's insideType-narrowing review skill — flags any escapes and suggests narrower typesLibrary evaluation skill — structured framework for picking a depRefactor playbooks — TS-specific patterns for safely moving codeTest writing skill — Vitest-flavored, behaviour-firstSeveral editorial skills for technical writing tasksCompare withAndrej Karpathy Skills — the single-file CLAUDE.md approach, broader and more philosophicalClaude Skills (Rezvani) — a 245+ skill bundle for breadth contrastFor the broader 2026 ranking question ("which AI tools and approaches are actually winning right now?"), see our today-dated speed ranking on AI image generators — the same instinct applies: small, opinionated, focused beats big and generic.

Anthropic's official open-source skill bundle for financial services hit GitHub on May 5, 2026 and pulled 9.3k stars in under 48 hours. The repo at anthropics/financial-services ships under Apache 2.0 and is the canonical reference for every Wall Street team standing up Claude-powered workflows in 2026. What is in the box: 10 named agent templates, 7 vertical skill bundles and 11 MCP data connectors, plus install scripts for Microsoft 365 (Excel, PowerPoint, Word) and a Claude Managed Agents cookbook for IT to deploy at scale. Three deployment surfaces — Claude Cowork plugins for desk users, Claude Code plugins for engineering teams, and Managed Agents for centralized rollout. Same code, different distribution. The 10 named agents cover the work analysts actually grind on: Pitch Agent (builds branded pitchbooks from comps and precedents), Meeting Prep Agent (client briefing packs), Earnings Reviewer (earnings calls and model updates), Model Builder (DCF, LBO and three-statement models in Excel), Market Researcher (industry overviews and competitive landscape), Valuation Reviewer (GP packages to LP reporting), GL Reconciler (general-ledger break detection and root-cause traces), Month-End Closer (accruals, roll-forwards, variance commentary), Statement Auditor (LP statement audits) and KYC Screener (document parsing plus rules engine). The 7 vertical skill bundles let you install the right scope for the right desk: financial-analysis (the core bundle — comps, DCF, LBO, three-statement models, deck QC, and all 11 data connectors), investment-banking (CIMs, teasers, merger models), equity-research (earnings notes, initiations, thesis tracking), private-equity (sourcing, screening, IC memos), wealth-management (client reviews, rebalancing, tax-loss harvesting), fund-admin (GL recon, accruals, NAV tie-out) and operations (KYC parsing and rules evaluation). The 11 MCP data connectors are the part that matters for production. Daloopa, Morningstar, S&P Global, FactSet, Moody's, MT Newswires, Aiera, LSEG, PitchBook, Chronograph and Egnyte all ship as first-class integrations on day one. Most of those vendors announced same-day support; Moody's launched its own MCP app the same morning. Anthropic did not stand this up alone — the data providers wired in alongside the launch, which tells you Wall Street's data stack has converged on MCP as the integration spec. The disclaimer woven through every skill is Anthropic's clearest signal that the team read the regulatory landscape before shipping. None of these agents auto-trade, auto-onboard customers or close books unsupervised. The agents draft analyst work product for human review and sign-off only. That is the right design for a regulated category — and it lines up with the Nature Human Behaviour finding that human-AI combinations on decision-making tasks need explicit verification steps to avoid degrading accuracy. Pairing notes: the README points to Claude Opus 4.7 as the recommended model. Opus 4.7 leads the Vals AI Finance Agent benchmark at 64.37%, ahead of every other publicly evaluated model. If you are running Sonnet for cost reasons, the agents work but expect lower first-pass accuracy on the modeling and audit tasks. Honest critique: this is reference code, not turnkey software. Installation is straightforward (the marketplace command is claude plugin marketplace add anthropics/claude-for-financial-services), but customizing the prompts, wiring the data connectors to your firm's authentication, and producing audit-ready output that survives compliance review is non-trivial. Budget engineering time for the integration, not just the install. And remember the productivity research: stacking all 10 agents on top of an existing Bloomberg + Excel + FactSet workflow puts you well into the brain-fry zone. Where it fits on Skila: the productivity context for these agents is in our analysis of why more AI tools makes your team slower. For document retrieval that pairs cleanly with the Earnings Reviewer and Model Builder agents, see PageIndex and PageIndex MCP. For governance across the resulting agent fleet, see Microsoft Agent 365.