AI Repos & Tools for Startup Founders
Open-source AI infrastructure, business automation tools, and startup-ready AI repositories.
Repositories
openclaw/openclaw
Your own personal AI assistant. Any OS. Any Platform. The lobster way. 🦞
310,753 starsollama/ollama
Ollama is an open-source platform written in Go that makes running large language models locally as straightforward as a single terminal command. Where tools like llama.cpp expose the raw inference engine, Ollama wraps the entire lifecycle -- model discovery, download, weight management, GPU acceleration, and serving -- into a polished developer experience. Running a model is as simple as typing `ollama run deepseek-v4` or `ollama run qwen3-coder`, and the system handles everything from pulling the right quantization for your hardware to allocating GPU memory and launching an API server. With over 164,000 GitHub stars and 14,700+ forks, Ollama has become the default way developers interact with open-source language models on their own machines. The project builds on llama.cpp for its inference backend but adds critical infrastructure layers on top: a model registry with thousands of pre-packaged models, automatic hardware detection across NVIDIA CUDA, AMD ROCm, and Apple Metal backends, and a REST API server that runs on localhost:11434 by default. The API is compatible with both the OpenAI Chat Completions format and, as of v0.14.0, the Anthropic Messages API -- meaning tools like Claude Code, Codex, Droid, and OpenCode can connect directly to local Ollama instances without proxy layers. The model library is one of Ollama's strongest differentiators. It provides ready-to-run versions of DeepSeek, Qwen, Gemma, Kimi-K2.5, GLM-5, MiniMax, gpt-oss, Mistral, LLaMA, Phi, and dozens more families across a range of parameter sizes and quantization levels. As of early 2026, the library supports over 40,000 model integrations. Specialized models like GLM-OCR for document understanding and Qwen3-VL for vision tasks are available alongside general-purpose chat and coding models. The `ollama launch` command, introduced in v0.15, streamlines the setup of coding agents by automatically configuring environment variables and connecting your preferred development tool to a local or cloud-hosted model. Ollama runs cross-platform on macOS, Linux, and Windows, with official Docker images for containerized deployments. Installation is a one-liner on every platform: a shell script on Linux, a DMG on macOS, or a PowerShell command on Windows. On Apple Silicon, Metal acceleration is automatic with no driver installation required -- the unified memory architecture means your full system RAM is available as GPU memory. On NVIDIA systems, CUDA drivers 535+ are detected automatically. AMD GPU support is available through ROCm 6.0+ on Linux. Recent releases have added structured output support (constraining model responses to JSON schemas), a built-in web search API, NVFP4 and FP8 quantization for up to 35 percent faster token generation on supported hardware, and a redesigned desktop application with file drag-and-drop for document reasoning. The v0.17.6 release in March 2026 refined tool calling for Qwen 3.5 models and fixed GLM-OCR prompt rendering. The project also offers cloud-hosted inference for larger models like GLM-4.6 and Qwen3-coder-480B that exceed typical consumer hardware budgets. Ollama's ecosystem integration is vast. Over 100 third-party projects connect to it, spanning web UIs (Open WebUI, LibreChat), desktop applications (AnythingLLM, Dify, Jan), orchestration frameworks (LangChain, LlamaIndex, Spring AI, Semantic Kernel), and automation platforms (n8n). Native client libraries are available in Python and JavaScript, with community libraries covering Go, Rust, Java, and more.
164,982 starslanggenius/dify
Dify is an open-source platform for building production-grade AI applications through a visual drag-and-drop workflow builder. Instead of writing boilerplate code to chain LLM calls, manage prompts, and wire up retrieval pipelines, developers lay out their logic on a canvas -- connecting model nodes, tool calls, conditional branches, and human-in-the-loop checkpoints into executable workflows. The result is a system that can go from prototype to production deployment without rewriting the orchestration layer. The platform supports hundreds of LLM providers out of the box: OpenAI GPT models, Anthropic Claude, Mistral, Llama 3, Qwen, and any provider exposing an OpenAI-compatible API. Switching between models is a dropdown change, not a code refactor. This provider-agnostic design means teams can start with a cloud API, benchmark against alternatives, and migrate to self-hosted models without touching their workflow logic. Dify ships with a full RAG pipeline built in. Upload PDFs, presentations, or plain text, and the system handles chunking, embedding, vector storage, and retrieval. Version 1.12.0 introduced Summary Index, which attaches AI-generated summaries to document chunks so semantically related content clusters together during retrieval. Version 1.13.0 added multimodal retrieval that unifies text and images into a single semantic space for vision-enabled reasoning. The agent capabilities layer supports both function-calling and ReAct-style reasoning with over 50 built-in tools including Google Search, DALL-E, Stable Diffusion, and WolframAlpha. Since v1.0.0, all models and tools have been migrated to a plugin architecture, so extending Dify with custom integrations no longer requires forking the core codebase. On the operational side, Dify includes a Prompt IDE for comparing model outputs side-by-side, LLMOps-grade logging and performance monitoring, and a Backend-as-a-Service API layer that lets frontend applications consume workflows through REST endpoints. The v1.13.0 release added a Human Input node that pauses workflows for human review, enabling approval gates and content moderation loops directly within automated pipelines. Deployment is flexible: Docker Compose for quick self-hosted setups (minimum 2 CPU cores, 4 GB RAM), Kubernetes via five community-maintained Helm charts, Terraform templates for Azure and Google Cloud, and AWS CDK for infrastructure-as-code deployments. Dify Cloud offers a managed option with 200 free GPT-4 calls to get started. Enterprise customers including Volvo Cars and Kakaku.com run Dify in production -- Kakaku.com reported 75% employee adoption with nearly 950 internal AI applications built on the platform.
132,700 starslangchain-ai/langchain
LangChain is the most widely used open-source framework for building LLM-powered applications and autonomous agents. It provides a standardized, composable interface across model providers — OpenAI, Anthropic, Google, Mistral, and 50+ others — so you can swap models without rewriting your logic. With 130,000+ GitHub stars and over 277,000 dependent projects, it's the de-facto standard for production RAG pipelines, multi-agent systems, and agentic workflows.
129,548 starsopen-webui/open-webui
Open WebUI is the most popular self-hosted AI chat platform on GitHub, with 127,000+ stars. It runs entirely offline and connects to Ollama, OpenAI-compatible APIs, and dozens of other LLM backends — giving you a ChatGPT-like experience without sending your data to any cloud.
127,076 starsanomalyco/opencode
118,000 stars in under a year. That's not a typo. OpenCode went from zero to the most-starred AI coding agent on GitHub faster than any developer tool in memory -- and it did it by being the one thing Cursor and Claude Code refuse to be: completely open source, completely provider-agnostic, and completely free. Here's the pitch that mass-converted developers: bring your own model, bring your own keys, keep your own data. OpenCode doesn't care if you're running Claude Opus, GPT-4.1, Gemini 2.5, or a local Llama instance through Ollama. It treats every provider as a first-class citizen. If you've ever felt locked into Anthropic's pricing because Claude Code only works with Claude, or locked into Cursor's $20/month because switching means losing your workflow -- OpenCode is the exit door 5 million developers already walked through. The architecture is what makes it stick. OpenCode runs a client/server split -- the AI agent runs as a background server while the TUI, desktop app, or IDE extension connects as a client. That means you can run the agent on a beefy remote machine and code from a thin laptop over SSH. Try doing that with Cursor. Two built-in agents handle different workflows: a build agent for writing and modifying code, and a read-only plan agent for exploring codebases without accidentally changing anything. There's also a general subagent that handles complex multi-step searches. LSP integration gives it real code intelligence -- not just pattern matching, but actual type-aware navigation and diagnostics. The release velocity tells its own story: 731 releases, 10,045 commits, v1.2.21 shipped March 7, 2026. The team at Anomaly (the same folks behind terminal.shop) ships daily. MCP support means you can extend it with the same server ecosystem Claude Code uses. Install takes one curl command. You're coding in 30 seconds.
121,683 starsggml-org/llama.cpp
llama.cpp is a high-performance C/C++ implementation for large language model inference, originally created by Georgi Gerganov in March 2023 shortly after Meta released its LLaMA models. With over 96,000 GitHub stars and more than 15,000 forks, it stands as one of the most consequential open-source AI projects ever built. The core premise is deceptively simple: strip away the Python runtime overhead and GPU requirements that dominate the LLM ecosystem, and rewrite inference from scratch in portable C/C++ with zero external dependencies. The project supports an extraordinary range of hardware backends -- eleven and counting -- including Apple Metal, NVIDIA CUDA, AMD HIP, Intel SYCL, Vulkan for cross-vendor GPU access, Ascend NPU via CANN, OpenCL for Adreno mobile GPUs, and experimental WebGPU for browser-based inference. This backend diversity means llama.cpp runs on everything from high-end data center GPUs to Raspberry Pis, Android phones, and iOS devices. A CPU+GPU hybrid inference mode allows models that exceed available VRAM to spill over to system RAM, making it practical to run 30B+ parameter models on consumer hardware. Quantization is where llama.cpp truly differentiates itself. The GGUF file format, which llama.cpp pioneered, supports quantization levels from 1.5-bit to 8-bit integers alongside standard float32, float16, and bfloat16 formats. Aggressive quantization (Q2_K through Q4_K) can reduce memory requirements by up to 75 percent, enabling models like LLaMA 2 13B or Mixtral 8x7B to run on machines with as little as 6-8 GB of RAM. The GGUF format has become a de facto standard, with Hugging Face providing native GGUF support and dedicated tools like GGUF-my-repo for model conversion. Beyond raw inference, llama.cpp ships with a production-ready HTTP server (llama-server) that exposes an OpenAI-compatible API. This means existing applications built against the OpenAI API can be pointed at a local llama.cpp server with minimal code changes. The server supports advanced features including speculative decoding for 1.5-2x throughput improvements on structured prompts, grammar-constrained output for reliable structured generation and function calling, and multimodal inference for vision-language models like LLaVA, MiniCPM, and Qwen2-VL. The project also includes VS Code and Vim/Neovim plugins for local code completion, a built-in web UI for interactive chat, and RPC-based distributed inference across multiple machines. Model compatibility spans over 50 text-only architectures (LLaMA, Mistral, Qwen, Phi, Gemma, Mamba, and many more) and 10+ vision-language models. Active development continues at a rapid pace, with over 8,200 commits and 758 open pull requests as of early 2026, including recent work on WebGPU shader optimization, CDNA3 tensor core flash attention for AMD MI300X GPUs, and native MXFP4 format support.
97,843 starsbrowser-use/browser-use
Browser Use is the leading open-source Python framework that enables large language models to autonomously control web browsers. With nearly 80,000 GitHub stars, it has become the de facto standard for AI-powered browser automation, allowing developers to build agents that can navigate websites, fill forms, extract data, and complete multi-step web tasks using natural language instructions. The framework takes a hybrid approach to web understanding, combining DOM parsing with vision-based analysis through Playwright. When an agent receives a task, Browser Use loads the target page, processes the DOM to identify interactive elements like buttons, links, and input fields, and optionally engages a vision model for layout-dependent decisions. The LLM then reasons about what actions to take and executes them in sequence. Browser Use is LLM-agnostic, supporting OpenAI GPT-4, Anthropic Claude, Google Gemini, and local models via Ollama, alongside its own optimized ChatBrowserUse model. The framework includes a CLI for interactive browser control, a template system for quick project scaffolding, and extensible custom actions via Python decorators. For production deployments, Browser Use offers a cloud browser infrastructure with stealth capabilities, CAPTCHA handling, and proxy rotation. Backed by a $17 million seed round led by Felicis with participation from Paul Graham and Y Combinator, Browser Use has achieved an 89.1% success rate on the WebVoyager benchmark across 586 diverse web tasks. Its MIT license and active development — with 8,739 commits and 118 releases — make it a robust foundation for any AI browser automation project.
80,692 starsinfiniflow/ragflow
RAGFlow is a leading open-source Retrieval-Augmented Generation engine that fuses deep document understanding with agentic AI capabilities to build a superior context layer for large language models. Unlike general-purpose RAG frameworks, RAGFlow specializes in extracting structured knowledge from complex, visually rich documents — including PDFs with tables, multi-column layouts, images, scanned copies, spreadsheets, slides, and web pages — with high fidelity. The platform provides template-based intelligent chunking with visual customization, high-precision hybrid search combining vector search, BM25, and custom scoring with advanced re-ranking, and grounded citations that reduce hallucinations by linking every answer back to traceable source references. RAGFlow includes a visual workflow builder for designing agentic RAG pipelines with memory support, Model Context Protocol (MCP) integration, and multi-modal model support for processing images within documents. It ships with Docker-based deployment in both lightweight (2 GB) and full-featured (9 GB) configurations, supports Elasticsearch and Infinity as storage backends, and works with configurable LLMs and embedding models. With 74,000+ GitHub stars and an Apache 2.0 license, RAGFlow has become one of the most popular open-source RAG solutions, particularly for enterprise use cases in equity research, legal analysis, and manufacturing where document intelligence is critical.
74,955 starsmsitarzewski/agency-agents
Agency-Agents is a production-ready collection of 144+ specialized AI agent personas organized across 12 divisions — Engineering, Design, Paid Media, Sales, Marketing, Product, Project Management, Testing, Support, Spatial Computing, Game Development, and Specialized. Each agent is a structured Markdown file that gives any LLM a specific professional identity, complete with domain expertise, a distinct communication style, battle-tested workflows, concrete deliverables, and measurable success metrics. Rather than relying on generic prompting, you activate a focused expert — a Frontend Wizard, Security Engineer, Brand Guardian, UX Researcher, or one of 140+ others — and the LLM narrows its context accordingly, reducing hallucinations and enforcing domain best practices. The project was born from a Reddit discussion about AI agent specialization and grew through months of community iteration into one of the fastest-starred repositories on GitHub. It integrates natively with Claude Code's /agents system by placing Markdown files in ~/.claude/agents/, and ships automated install scripts that convert agents for Cursor, Aider, Windsurf, Gemini CLI, OpenCode, and more. With 43.9K stars and 6.6K forks as of March 2026, agency-agents has become the de facto starting point for teams that want to run structured multi-agent workflows from their IDE.
43,880 starsKeygraphHQ/shannon
Shannon Lite is a fully autonomous AI pentester that does something most security tools only promise: it actually exploits the vulnerabilities it finds. Point it at a running web app with source code access, and Shannon analyzes the codebase to identify attack vectors, then uses browser automation and command-line tools to execute real injection attacks, authentication bypasses, SSRF, XSS, and more -- all without human intervention. The numbers back it up. On a cleaned, hint-free variant of the XBOW benchmark (104 intentionally vulnerable apps), Shannon scored 96.15%, successfully executing 100 out of 104 exploits. The "hint-free" part matters: the benchmark was stripped of descriptive variable names, comments, and filenames that could artificially inflate results. In real-world testing against OWASP Juice Shop, it uncovered 20+ critical vulnerabilities including full auth bypass and database exfiltration via injection. The workflow is refreshingly simple. Clone the repo, drop your target's source code into the repos directory, and run a single command: `./shannon start URL=https://your-app.com REPO=repo-name`. Shannon handles everything from 2FA/TOTP login flows to parallel vulnerability scanning across attack categories to final report generation. It only reports proven, exploitable findings with copy-paste proof-of-concept code -- no theoretical warnings or false positive noise. Under the hood, Shannon integrates Nmap, Subfinder, WhatWeb, and Schemathesis for comprehensive reconnaissance and API testing. It runs in Docker containers and supports Anthropic Claude, AWS Bedrock, and Google Vertex AI as LLM backends. A full pentest run typically takes 1 to 1.5 hours and costs roughly $50 in API calls using Claude 4.5 Sonnet. Shannon Pro (commercial) extends Lite with SAST, SCA, and secrets scanning, correlating static analysis findings with dynamic exploit validation in a single workflow. But Lite alone is a serious tool for security-conscious teams who want automated, proof-based vulnerability assessment.
33,465 starskhoj-ai/khoj
Khoj is an open-source, self-hostable AI second brain that turns your personal documents into a searchable, conversational knowledge base. It uses retrieval-augmented generation (RAG) to answer questions by combining your local files — PDFs, Markdown, Org mode, and more — with real-time web search results. You can run it entirely on your own hardware for full data privacy, or connect it to cloud LLM providers like OpenAI, Anthropic, or Google. What makes Khoj worth the 33,000+ stars on GitHub is its flexibility. It supports over 20 LLM backends including Llama 3, GPT, Claude, Gemini, DeepSeek, Qwen, and Mistral. You access it through whatever interface fits your workflow — browser, Obsidian plugin, Emacs client, desktop app, mobile, or WhatsApp. Custom agents let you build specialized assistants with their own knowledge bases and tool access. Scheduled automations handle recurring research tasks and push notifications when results are ready. Recent additions include Pipali, an open-source desktop AI co-worker, and a v2.0 beta with significant architecture improvements. The project ships 170+ releases under AGPL-3.0 with an active community on Discord.
33,378 starsMCP Servers
Context7
Context7 is the most popular MCP server for injecting up-to-date, version-specific documentation and code examples directly into AI prompts. Built by Upstash, it solves the critical problem of LLMs generating broken code from outdated training data by pulling the latest official documentation from over 33,000 indexed libraries at the moment of prompting. Context7 uses an Upstash Vector Database with the DiskANN algorithm to perform semantic search across library documentation, returning relevant code snippets and API references filtered by topic. It exposes two MCP tools: resolve-library-id (which converts human-readable library names into Context7-compatible identifiers using LLM-powered search and ranking) and get-library-docs (which retrieves current documentation sections and code examples for a specific library and version). With over 44,000 GitHub stars and 240,000 weekly npm downloads, Context7 is the number one MCP server by adoption. Integration is seamless across all major AI code editors including Cursor, Claude Desktop, Windsurf, VS Code, and Claude Code. Users simply add 'use context7' to any prompt to activate real-time documentation injection, eliminating hallucinated APIs and outdated code generation.
48,933 starsNotion MCP Server
Official first-party MCP server from Notion Labs that gives AI assistants full read and write access to Notion workspaces — create pages, query databases, manage comments, and search across connected tools including Slack, Google Drive, and Jira.
4,038 starsAzure MCP Server
Microsoft's official Model Context Protocol server for Azure. Gives AI agents natural-language access to 40+ Azure services through 200+ structured tools — from storage and databases to AI/ML services, Kubernetes, and Cosmos DB. Works with any MCP client including Claude, GitHub Copilot, VS Code, and OpenAI agents.
2,773 starsSupabase
The Supabase MCP Server connects your Supabase projects to AI assistants like Claude, Cursor, and Windsurf through the Model Context Protocol standard. It exposes 34 tools across eight feature groups — account management, documentation search, database operations, debugging, development utilities, Edge Functions, branching, and storage — giving LLMs direct access to query databases, apply migrations, deploy Edge Functions, manage branches, and monitor logs. Built by the official Supabase community organization, this server supports both cloud-hosted and local deployment modes. The cloud version at mcp.supabase.com uses OAuth 2.1 for seamless authentication, while local installations work through personal access tokens. Configuration is granular: developers can enable read-only mode to prevent write operations, scope the server to a single project to limit blast radius, and selectively enable or disable feature groups to minimize the exposed tool surface. Security is a first-class concern. SQL query results are wrapped with prompt injection mitigations, and the server recommends connecting to development projects rather than production databases. The branching system allows safe experimentation by creating isolated database copies before merging changes to production. Currently in public alpha with the npm package @supabase/mcp-server-supabase, the server sees active development with approximately 50,000 total npm downloads and growing adoption across the AI-assisted development ecosystem.
2,533 starsStripe Agent Toolkit
Ask Claude to create a payment link, email the invoice, and apply a 20% coupon -- in one prompt. The Stripe Agent Toolkit MCP server turns your AI assistant into a billing operations co-pilot that can read, write, and manage your entire Stripe account without you ever opening the dashboard. The server exposes 30+ tools covering the full billing lifecycle. Create and list customers, build product catalogs with tiered pricing, generate payment links for instant checkout, draft and finalize invoices with line items, manage subscriptions (create, update, cancel), process refunds, handle disputes, and create coupon codes. Utility tools like search_stripe_resources and fetch_stripe_resources let Claude query any Stripe object by ID or search term, and search_stripe_documentation gives it access to Stripe's knowledge base for answering integration questions on the fly. Stripe hosts a remote MCP server at mcp.stripe.com with OAuth authentication, so you can connect from Claude Desktop, Cursor, or VS Code without managing API keys locally. For local development or CI pipelines, run `npx -y @stripe/mcp --api-key=sk_test_...` and you are up in under a minute. Security is handled through Restricted API Keys -- you decide exactly which Stripe resources your AI assistant can touch, down to individual API methods. The real power shows up when you combine Stripe MCP with other servers. Pair it with GitHub MCP to build an end-to-end workflow: customer reports a billing bug in an issue, Claude reads the issue, queries Stripe for the customer's invoice history, identifies the problem, issues a refund, and posts the resolution back to the issue. Or connect it with a database MCP server to reconcile your internal records against Stripe's data. Beyond MCP, the toolkit also works as a standalone SDK for OpenAI Agents, LangChain, CrewAI, and Vercel AI SDK -- but the MCP server is the zero-config path for Claude users. This is an official Stripe product built on the Stripe Node SDK, backed by Stripe's security team, and MIT-licensed. If you run any kind of SaaS, e-commerce, or subscription business and use an AI coding assistant, this server pays for itself the first time you resolve a billing support ticket without leaving your editor.
1,359 starsMongoDB MCP Server
Ask Claude to query your MongoDB collections, build aggregation pipelines, and manage Atlas clusters -- all from a single prompt. The MongoDB MCP Server turns your AI assistant into a full-fledged database operator that reads, writes, and analyzes your data without you ever opening mongosh or Compass. The server exposes over 20 database tools covering the complete MongoDB workflow: run find and aggregate queries, insert and update documents, create and drop collections, manage indexes, and inspect collection schemas and statistics. On the Atlas side, it handles cluster provisioning, database user management, IP access list configuration, and performance advisor recommendations. Need to figure out why a query is slow? Ask Claude to pull the performance advisor data and suggest index optimizations -- it has direct access. A built-in read-only mode makes this safe for production databases. Lock it down with a single flag and let Claude analyze your data without any risk of accidental writes or deletes. The server also includes a MongoDB documentation knowledge base, so Claude can reference official docs when answering questions about query syntax or schema design patterns. Setup takes under 60 seconds: run npx mongodb-mcp-server with your connection string, point your MCP client at it, and start querying. Docker support is available for containerized environments. HTTP transport mode opens the door to remote access and multi-client setups. Pair this with the GitHub MCP Server for a pipeline where Claude finds a data bug in your app code, queries production to confirm it, and opens a PR with the fix. Backend engineers, data analysts, and DevOps teams managing Atlas deployments will get the most value here.
958 starsBrave Search
The Brave Search MCP Server is the official Model Context Protocol integration from Brave Software that connects AI assistants to Brave's independent web search index. Unlike most search APIs that rely on Bing or Google indexes, Brave operates its own large-scale web index built with contributions from its privacy-preserving Web Discovery Project. The server exposes six specialized MCP tools: brave_web_search for comprehensive web searches with advanced filtering and custom re-ranking via Goggles, brave_local_search for discovering nearby businesses with ratings, hours, and contact details, brave_image_search and brave_video_search for rich media discovery with safety filtering and metadata, brave_news_search for current news articles with freshness controls, and brave_summarizer for AI-generated summaries from search results using Brave's summarization API. The server supports both STDIO and HTTP transport modes, making it compatible with all major MCP clients including Claude Desktop, Claude Code, Cursor, Windsurf, and VS Code. Configuration is straightforward with a single BRAVE_API_KEY environment variable, and the free tier provides 2,000 queries per month at no cost. Version 2.x of the server returns response objects that closely mirror the original Brave Search API structure, providing rich metadata including spell check suggestions, content safety classifications, and location-aware results. The package replaces the now-deprecated @modelcontextprotocol/server-brave-search and is actively maintained by Brave Software under the MIT license.
774 starsClickHouse MCP Server
Ask Claude to analyze a billion rows in your ClickHouse cluster and get an answer in under a second. That is the reality once you wire up the official ClickHouse MCP server. It turns any MCP-compatible AI assistant into a direct interface to your columnar database -- no intermediary scripts, no copy-pasting SQL output into chat windows, no context lost between questions. The server exposes three core tools to the AI: run_query for executing SQL, list_databases for discovering what clusters contain, and list_tables for browsing schemas with pagination and optional column metadata. By default, every query runs in read-only mode, so there is zero risk of an AI accidentally mutating production data. If you explicitly need write access for ETL or data pipeline work, you opt in with environment flags -- and destructive operations like DROP require a second flag on top of that. The safety model is well thought out. What surprised me most during setup is the chDB integration. With a single environment variable (CHDB_ENABLED=true), you unlock an embedded ClickHouse engine that can query Parquet files, CSVs, and remote URLs directly -- no data loading step required. Point Claude at a 500MB Parquet file on S3 and ask questions about it. That alone is worth the install. Setup takes about five minutes. Install via pip or uv, drop the config block into your Claude Desktop settings (or any MCP client config), add your ClickHouse credentials as environment variables, and restart. The server supports stdio, HTTP, and SSE transports, with bearer token authentication baked in for production HTTP deployments. There is even a /health endpoint for monitoring. This pairs naturally with other MCP servers in a data engineering stack. Combine it with the GitHub MCP server for code-aware analytics workflows, or the Brave Search server to let Claude cross-reference your internal metrics against public benchmarks. ClickHouse is already powering analytics at Cloudflare, Uber, and eBay -- now your AI assistant speaks the same language those systems do.
707 starsPlanetScale
The PlanetScale MCP Server is an official, hosted Model Context Protocol server that bridges AI assistants and PlanetScale MySQL databases. Unlike self-hosted MCP servers that require local installation and credential management, PlanetScale runs its server at mcp.pscale.dev, handling authentication through OAuth with granular permission scopes. This architecture eliminates the need to store database credentials on developer machines entirely. The server exposes 16 tools organized into four functional groups. Infrastructure tools let AI agents list organizations, databases, branches, regions, and cluster SKUs. Query tools handle both read and write operations with built-in safety rails: read queries automatically route to replicas when available, reducing primary database load, while write queries enforce destructive query protection that blocks UPDATE and DELETE statements lacking a WHERE clause and prevents TRUNCATE operations altogether. Schema management tools provide branch schema inspection and performance-oriented schema recommendations powered by PlanetScale Insights data. A documentation search tool rounds out the toolset by letting agents query PlanetScale docs directly. Security is deeply integrated into the server's design through three mechanisms. First, all database credentials are ephemeral, created on demand for each query execution and deleted immediately after. Second, every query is tagged with a source=planetscale-mcp SQL comment, making all AI-originated queries visible in PlanetScale Insights for auditing and performance monitoring. Third, DDL operations such as CREATE TABLE, ALTER TABLE, and DROP TABLE require explicit human confirmation before the server executes them, preventing accidental schema changes from autonomous AI agents. The server supports a wide range of MCP clients including Claude Desktop, Claude Code, Cursor, VS Code, Notion, Zed, and CLI tools like OpenCode, Codex CLI, Gemini CLI, and Amp CLI. Quick-install options are available for Cursor and VS Code through one-click buttons, while Claude Desktop users can add it through the Connectors Directory. Launched in January 2026, the PlanetScale MCP Server represents a production-grade approach to database-AI integration with enterprise security guarantees.
647 starsNeon MCP Server
Tell Claude 'create a new Postgres database for my side project' and watch it happen in 3 seconds — no psql, no AWS console, no connection string hunting. The Neon MCP Server turns Claude into a full database administrator for Neon's serverless Postgres platform. You're prototyping at 1am and need a fresh database branch to test a migration. Instead of clicking through dashboards, you type 'create a branch from production called test-auth-migration.' Claude calls Neon's API, the branch exists instantly with a full copy of your production data, and you get the connection string back in the same conversation. The real power isn't just CRUD operations — it's how branching changes your development workflow. Neon branches are copy-on-write, meaning Claude can spin up an isolated database copy in under a second without duplicating storage. Ask Claude to 'run this migration on a test branch, verify it works, then show me the schema diff.' That's a three-step pipeline that used to require 15 minutes of manual context-switching. The server exposes every Neon Management API operation: create and manage projects, databases, branches, roles, and endpoints. Run SQL queries directly from Claude. Monitor database activity and resource utilization. And the newest addition — the provision_neon_auth tool now returns your authentication configuration idempotently, so you can retrieve JWKS URLs without separate API calls. Setup is one command — configure your Neon API key, add the server to your MCP config, and you're managing databases conversationally. Pair it with the GitHub MCP server and you've got Claude managing your database migrations, committing schema changes, and opening PRs — all from natural language. The caveat: this is designed for development and testing, not production database management. Don't point it at your production cluster and ask Claude to 'clean up unused tables' at 3am. For dev workflows, though, it eliminates an entire category of context-switching that drains focus.
560 starsChroma MCP Server
Give Claude direct access to a vector database and suddenly every conversation has long-term memory. The Chroma MCP Server connects Claude Desktop (or any MCP-compatible client) to Chroma, the open-source embedding database, letting you store documents, run semantic searches, and manage collections without leaving your chat window. The standout capability here is semantic document search. Instead of relying on exact keyword matches, you can ask Claude to find documents by meaning. Store your engineering docs, support tickets, or research papers in a Chroma collection, then query them with natural language. Claude handles the embedding and retrieval behind the scenes using any of six supported embedding providers: the built-in default model, OpenAI, Cohere, Jina, Voyage AI, or Roboflow. Setup takes under five minutes for the basic ephemeral mode. Add one JSON block to your Claude Desktop config, restart, and you have 12 tools available immediately. For anything beyond experimentation, switch to persistent mode to keep your data across restarts, or point it at a self-hosted Chroma instance or Chroma Cloud for team-wide access. The server exposes fine-grained control over collections and documents. You can create collections with custom HNSW index parameters, add documents in bulk with metadata, run filtered queries combining semantic similarity with metadata constraints, and paginate through large result sets. Batch operations make it practical to index hundreds or thousands of documents in a single session. One thing worth noting: this is the official server built by the Chroma team (chroma-core on GitHub), not the community-built chroma-mcp-server package. It runs via uvx with zero dependency management on your part. Python 3.10+ is the only prerequisite. The server supports environment variables and dotenv files for credential management, so API keys for cloud deployments and embedding providers stay out of your command-line history.
515 starsCloudflare MCP Server
Ask Claude to deploy a Worker, create a KV namespace, configure DNS records, and set up an R2 bucket -- in one prompt. That's not a demo. That's what 2,500+ Cloudflare API endpoints look like when they're accessible through two MCP tools instead of 2,594. Here's the trick that makes this server different from every other API-wrapper MCP: it doesn't dump the entire API spec into your context window. Traditional MCP servers expose one tool per API endpoint -- Cloudflare has 2,594 endpoints, which would eat 244,000 tokens just to describe. This server uses Code Mode: it gives Claude a search() tool to query the API specification and an execute() tool to call endpoints. The entire server description fits in 1,069 tokens. That's 0.5% of what the brute-force approach would cost. You're managing a Cloudflare account. You need to check which Workers are deployed, read the KV store for a specific key, find the DNS records for a domain, and check R2 storage usage. Without this server, that's four separate API calls you'd have to construct manually, checking docs for each endpoint. With it, you say: 'Show me all deployed Workers, the value of key "config" in my production KV, DNS records for example.com, and R2 usage this month.' Claude searches the spec, finds the four endpoints, and executes them. Setup is instant -- it's a remote MCP server hosted at mcp.cloudflare.com. No npm install, no Docker, no local dependencies. Add one URL to your MCP config, authenticate with OAuth or an API token, and you're managing infrastructure from Claude. The services you get access to cover everything Cloudflare offers: Workers, KV, R2, D1, Pages, DNS, Firewall, Load Balancers, Stream, Images, AI Gateway, Vectorize, Access, Gateway, and GraphQL Analytics. Every service Cloudflare has ever shipped is available through this single MCP server. Pair this with the GitHub MCP server and you've got a pipeline: Claude reads your repo, deploys to Workers, configures the CDN, and monitors performance -- all from a single conversation.
263 stars