Editor's picks

One desktop agent that runs Claude Code and Codex with shared memory and skills

Why: Reuses a Claude Code or Codex plan you already pay for, so you get a desktop agent without a second model bill.

Gained about 4.1k stars in the last 7 days (3.3k → 7.4k); last updated 10-27

Anthropic's role-based plugins for Claude Cowork

Why: Ready-made starting points for many business roles, each bundling skills, connectors, commands and sub-agents.

Gained about 1.9k stars in the last 7 days (31.5k → 33.4k); last updated 10-27

Agent meta-harness for Claude Code and Codex swarms

Why: Multi-agent coordination and cross-session memory without changing how you use Claude Code day to day.

Gained about 1.4k stars in the last 7 days (78.4k → 79.8k); last updated 10-27

Give a coding agent a local video editor that checks its own output

Why: The probe-edit-verify workflow makes the agent check its own output instead of guessing FFmpeg flags and hoping.

Gained about 1.1k stars in the last 7 days (2.0k → 3.1k); last updated 10-26

Hosted MCP gateway to thousands of ready-made Apify scrapers

Why: Access to a very large catalogue of ready-made scrapers instead of a single tool.

Gained about 900 stars in the last 7 days (12.6k → 13.5k); last updated 10-27

An operating layer for Hermes Agent that adds workflows, model routing and honest evidence records without replacing it

Why: Adds structured workflows to Hermes while keeping Hermes and its native skills in place.

Gained about 800 stars in the last 7 days (3.6k → 4.4k); last updated 10-27

Keep raw tool output out of your agent's context window

Why: Large context savings on tool-heavy sessions by keeping raw output in a sandbox.

Gained about 700 stars in the last 7 days (31.8k → 32.5k); last updated 10-27

An open-source desktop office suite that edits real .docx, .xlsx and .pptx files with an AI whose changes you can review

Why: AI changes are reviewable as tracked changes and diffs, with per-turn rollback.

Gained about 600 stars in the last 7 days (10.5k → 11.1k); last updated 10-27

One Rust binary that indexes any folder so coding agents search instead of reading files

Why: Elasticsearch-compatible API means existing clients, dashboards and tooling work against it unchanged.

Gained about 4.4k stars in the last 7 days (12.4k → 16.8k); last updated 10-26

Google's open-source terminal AI agent with a free tier

Why: Generous free tier with a personal Google account, and no API key management needed to get started.

Gained about 1.0k stars in the last 7 days (107.5k → 108.5k); last updated 10-26

Builds a markdown knowledge graph of your codebase so coding agents explore less

Why: Rebuilds the structural graph against the working tree on every query, so it reflects uncommitted edits.

Gained about 300 stars in the last 7 days (10.5k → 10.8k); last updated 10-26

A lifelike LLM chat agent for QQ groups that talks casually, knows when to speak and learns the group's slang

Why: Learns group slang and each person's habits over time.

Gained about 300 stars in the last 7 days (6.3k → 6.6k); last updated 10-26

Convert Korean HWP/HWPX, PDF and Office files to Markdown, with CLI and MCP

Why: Handles Korean HWP and HWPX formats alongside PDF, Excel, Word, PowerPoint and images in one tool.

Gained about 200 stars in the last 7 days (2.6k → 2.8k); last updated 10-26

SDK for agent-native apps with generative UI

Why: Purpose-built front-end layer for agents: generative UI, shared state and human-in-the-loop out of the box.

Gained about 200 stars in the last 7 days (38.1k → 38.3k); last updated 10-26

Ultra-lightweight self-hosted personal AI agent framework

Why: Reaches you through a WebUI, terminal and many chat apps, including Feishu, WeChat and Linear.

Gained about 200 stars in the last 7 days (49.2k → 49.4k); last updated 10-26

Open-source, local AI presentation generator with API and MCP server

Why: Can create a template from an existing PPTX so new decks follow your brand design.

Gained about 200 stars in the last 7 days (11.2k → 11.4k); last updated 10-26

Real-time global news and risk dashboard with maps, AI briefs and an MCP server

Why: Can run fully locally with Ollama and starts with no environment variables or API keys.

Gained about 700 stars in the last 7 days (88.8k → 89.5k); last updated 10-25

Claude Code skills that run ML research loops with a second model as reviewer

Why: Cross-model design uses a second model as an adversarial reviewer instead of Claude reviewing itself.

Gained about 400 stars in the last 7 days (17.7k → 18.1k); last updated 10-24

MIT Rust MCP server and desktop Studio with 48 tools for building, validating, reasoning over and versioning OWL ontologies with Claude

Why: Terraform-style plan, apply, drift and lineage commands give change control over ontologies

Gained about 300 stars in the last 7 days (1.1k → 1.4k); last updated 10-23

Seven-stage Chinese learning map from LLM basics to building multi-agent systems

Why: Clear seven-stage path with required hands-on exercises (each with a goal and success criteria) rather than a flat link list.

Gained about 300 stars in the last 7 days (7.7k → 8.0k); last updated 10-24

Local-first AI video editor where agents edit a real multitrack timeline

Why: AI edits land on a real multitrack timeline, so results stay editable and undoable

Gained about 200 stars in the last 7 days (2.5k → 2.7k); last updated 10-21

Open-source proxy for LLM, MCP and A2A traffic with guardrails and observability

Why: One proxy for LLM calls (unified OpenAI-compatible API with spend controls, failover and load balancing), MCP tools and A2A agent traffic.

Gained about 200 stars in the last 7 days (5.4k → 5.6k); last updated 10-25

Secure cloud sandboxes for running AI-generated code

Why: Isolated sandboxes keep untrusted, AI-generated code away from your infrastructure.

Gained about 200 stars in the last 7 days (14.4k → 14.6k); last updated 10-24

MCP server that lets agents drive iOS and Android devices, simulators and emulators

Why: Prefers accessibility trees, so many interactions need no vision model, and falls back to screenshots when labels are missing.

Gained about 200 stars in the last 7 days (9.8k → 10.0k); last updated 10-25

A Claude Code skill that turns a technical idea into an interactive isometric machine drawn in SVG, WebGL or 3D

Why: Zero dependencies and no API keys; outputs plain HTML, SVG or React components you can drop into docs.

Gained about 4.2k stars in the last 7 days (15.9k → 20.1k); last updated 10-18

Free local video editor with captions, voice cloning and an MCP API

Why: No watermark, account or paywall, and everything runs locally and offline once models are downloaded.

Gained about 2.4k stars in the last 7 days (8.9k → 11.3k); last updated 10-24

Connect Blender to any LLM for prompt-driven 3D modelling

Why: Can pull assets and models from Poly Haven, Sketchfab and AI generators such as Hyper3D Rodin and Hunyuan3D.

Gained about 700 stars in the last 7 days (31.5k → 32.2k); last updated 10-24

MCP server that lets LLMs control the Windows desktop with mouse, keyboard and shell

Why: Gives any LLM control of the Windows desktop (mouse, keyboard, windows, apps) without a vision model or fine-tuning.

Gained about 700 stars in the last 7 days (9.5k → 10.2k); last updated 10-24

20 MB database client for 70+ databases with an AI assistant and MCP server

Why: Supports 70+ databases, including relational, NoSQL, vector and several Chinese databases, in one tool.

Gained about 600 stars in the last 7 days (27.1k → 27.7k); last updated 10-24

Python framework for role-based multi-agent workflows

Why: Two complementary models in one framework: autonomous Crews where flexibility helps, deterministic Flows where control matters.

Gained about 300 stars in the last 7 days (59.8k → 60.1k); last updated 10-24

Apache-2.0 self-hosted Community Edition that turns Codex, Claude Code and other agent harnesses into one OpenAI Responses-compatible API

Why: Handles persistent sessions, streaming progress, files, artifacts, cancellation and structured failures

Gained about 300 stars in the last 7 days (3.1k → 3.4k); last updated 10-24

Open course: build a Claude Code-like agent from scratch

Why: Clear, incremental structure: each lesson adds exactly one mechanism with runnable code.

Gained about 400 stars in the last 7 days (78.8k → 79.2k)

Adaptive Python web scraping framework with anti-bot fetchers

Why: Adaptive selectors that survive website redesigns reduce scraper maintenance.

Gained about 1.0k stars in the last 7 days (88.3k → 89.3k); last updated 10-22

Python framework for transparent, production-ready agents

Why: Multi-agent features (pipelines, team pipelines, A2A) plus realtime voice agents.

Gained about 400 stars in the last 7 days (33.4k → 33.8k); last updated 10-23

Curated list of MCP servers, organised by category

Why: Very broad coverage, organised into clear categories.

Gained about 300 stars in the last 7 days (96.5k → 96.8k); last updated 10-23

Desktop AI client for many LLM providers and assistants

Why: One app for many providers, including local models, with multi-model conversations side by side.

Gained about 200 stars in the last 7 days (52.7k → 52.9k); last updated 10-23

GitHub's official MCP server for repos, issues and PRs

Why: Official and actively maintained by GitHub, covering repositories, issues, pull requests, Actions, security alerts and more.

Gained about 200 stars in the last 7 days (33.7k → 33.9k); last updated 10-23

MCP bridge between AI clients and a live Godot editor

Why: Large tool set: 120+ operations including scene building, node editing, project search and running tests

Gained about 200 stars in the last 7 days (3.0k → 3.2k); last updated 10-23

Open-source deep research agent that writes cited reports

Why: Produces long, cited reports from many sources, with source tracking for each claim.

Gained about 200 stars in the last 7 days (30.1k → 30.3k); last updated 10-10

Durable workflow engine for microservices and AI agents

Why: Durable, resilient execution proven at large scale, with retries, cancellation and approvals.

Last updated 10-23

Self-hosted AI assistant you chat with in WhatsApp, Slack, Telegram

Why: Model-agnostic: works with hosted and local model providers, which can be swapped without changing the rest of the setup.

Gained about 600 stars in the last 7 days (392.2k → 392.8k); last updated 10-22

Code knowledge graph that cuts agent token use on codebases

Why: Large token savings on code exploration: one graph query replaces many search and read steps.

Gained about 500 stars in the last 7 days (46.7k → 47.2k); last updated 10-22

ByteDance's open-source super-agent runtime for long tasks

Why: Batteries included: sandbox, filesystem, memory, skills, sub-agents and scheduled tasks come out of the box instead of being assembled by hand.

Gained about 300 stars in the last 7 days (84.0k → 84.3k); last updated 10-22

Let coding agents debug and inspect a live Chrome

Why: Brings real DevTools capabilities to agents: performance traces with insights, network analysis and source-mapped console errors.

Gained about 300 stars in the last 7 days (53.7k → 54.0k); last updated 10-22

Agent operating system built around isolated workspaces with white-box memory

Why: White-box memory: you can see every memory entry, find the one that caused a mistake and edit it.

Gained about 300 stars in the last 7 days (4.5k → 4.8k); last updated 10-20

Self-hosted AI trading OS: strategies, backtests, paper and live execution

Why: Self-hosted and local-first, so strategy code, data and broker credentials stay with you.

Gained about 200 stars in the last 7 days (12.8k → 13.0k); last updated 10-22

Local-first 3D building editor you can drive from agents through MCP

Why: Includes an authenticated local MCP service and a `pascal mcp connect` command so agents can work on a model.

Gained about 200 stars in the last 7 days (24.9k → 25.1k); last updated 10-22

Open-source Evernote-style knowledge base with built-in MCP and free serverless hosting

Why: Data stays in SQLite with REST APIs and lossless ZIP export, so you are not locked in

Gained about 200 stars in the last 7 days (2.3k → 2.5k); last updated 10-22

Compress tool output and logs before they reach the LLM

Why: Content-aware compression for JSON, code and prose, with reversible retrieval of the originals.

Gained about 1.9k stars in the last 7 days (75.4k → 77.3k); last updated 10-22

Workflow automation platform for building AI agents

Why: AI features are first-class: multi-step agents, tool use, human approval steps and model switching without changing the workflow structure.

Gained about 800 stars in the last 7 days (207.5k → 208.3k); last updated 10-22

Assign issues to AI coding agents like teammates

Why: Strong oversight: replayable execution logs, per-run cost analytics and review gates before anything ships.

Gained about 600 stars in the last 7 days (52.9k → 53.5k); last updated 10-22

Up-to-date library docs for your AI coding assistant

Why: Directly reduces hallucinated APIs and outdated examples by grounding the agent in current, version-specific docs.

Gained about 400 stars in the last 7 days (63.1k → 63.5k); last updated 10-21

Open-source RAG engine with deep document parsing and citations

Why: Grounded answers with traceable citations, which helps reviewers trust and verify output.

Gained about 300 stars in the last 7 days (92.3k → 92.6k); last updated 10-21

Microsoft's official MCP server for browser automation

Why: Works from accessibility snapshots, so no vision model is needed and actions are more deterministic than screenshot-based clicking.

Gained about 300 stars in the last 7 days (38.2k → 38.5k); last updated 10-22

MIT self-hostable AI security operations center with a local model, auditable agent investigations and human-approved responses

Why: Every agent prompt and tool call is logged, and responses need human approval

Gained about 200 stars in the last 7 days (2.4k → 2.6k); last updated 10-22

High-performance AI gateway with one OpenAI-compatible API for 23+ providers

Why: Automatic fallbacks and load balancing across providers, models and API keys.

Gained about 200 stars in the last 7 days (8.8k → 9.0k); last updated 10-21

Agent meta-harness for Claude Code and Codex swarms

Why: Multi-agent coordination and cross-session memory without changing how you use Claude Code day to day.

Gained about 3.0k stars in the last 7 days (74.7k → 77.7k); last updated 10-20

Keep raw tool output out of your agent's context window

Why: Session state persists through context compaction without re-flooding the window.

Gained about 2.3k stars in the last 7 days (29.0k → 31.3k); last updated 10-20

Anthropic's role-based plugins for Claude Cowork

Why: Ready-made starting points for many business roles, each bundling skills, connectors, commands and sub-agents.

Gained about 1.1k stars in the last 7 days (29.7k → 30.8k); last updated 10-20

Hosted MCP gateway to thousands of ready-made Apify scrapers

Why: Hosted server means no local setup, and tools can be restricted to save context.

Gained about 1.1k stars in the last 7 days (11.2k → 12.3k); last updated 10-20

Sandboxed desktops and drivers for computer-use agents

Why: Works with any agent and model through CLI, MCP and SDKs in four languages.

Gained about 700 stars in the last 7 days (29.7k → 30.4k); last updated 10-20

Semantic code navigation and editing tools for coding agents

Why: Symbol-level tools (references, renames, moves) replace many error-prone text edits with single atomic operations.

Gained about 200 stars in the last 7 days (30.3k → 30.5k); last updated 10-20

Self-hosted backend for xiaozhi-esp32 voice hardware with MCP and knowledge base

Why: Supports MQTT+UDP, WebSocket, MCP endpoints, voiceprint recognition and a knowledge base.

Gained about 200 stars in the last 7 days (10.8k → 11.0k); last updated 10-20

Trend and public-opinion monitor for Chinese platforms

Why: Keyword filtering and AI briefings cut through information overload.

Gained about 200 stars in the last 7 days (62.9k → 63.1k); last updated 10-20

Debloat any Android or Google TV over Wi-Fi into an Apple TV-style home

Why: Tiny 300 KB native launcher is lightning-fast and responsive.

Viral breakout gaining ~7.5k stars in the last 7 days (0 → 7.5k); last updated 10-20

MIT npm tool that keeps your coding agent's lessons in the repo and maps the codebase for Claude Code and others; benchmark claims are the project's own

Why: Graft builds a readable graph of linked markdown files with no embeddings or server, wired into Claude Code and other agents

Gained about 500 stars in the last 7 days (10.0k → 10.5k); last updated 10-19

Control the Unity Editor from AI assistants via MCP

Why: Broad editor control: scenes, GameObjects, scripts, assets, tests, profiling and builds.

Gained about 300 stars in the last 7 days (14.9k → 15.2k); last updated 10-19

Single-binary Rust MCP server and native KiCAD 10 plugin with 185 tools for AI-assisted PCB design

Why: On-demand toolset router keeps initial MCP context overhead around ~2K tokens instead of ~23K tokens.

Gained about 338 stars in the last 7 days (962 → 1.3k); last updated 10-19

Ultra-lightweight self-hosted personal AI agent framework

Why: Reaches you through a WebUI, terminal and many chat apps, including Feishu, WeChat and Linear.

Gained about 200 stars in the last 7 days (49.0k → 49.2k); last updated 10-19

MCP server for "vibe reversing" in IDA Pro

Why: Lets an assistant decompile, disassemble, trace cross-references and read memory in a live IDA database.

Gained about 200 stars in the last 7 days (12.8k → 13.0k)

Desktop app for working with AI agents on your own files, powered by OpenCode

Why: Permission prompts (allow once, always or deny) and debug exports make agent actions easier to audit.

Gained about 200 stars in the last 7 days (6.9k → 7.1k); last updated 10-19

Google's open-source terminal AI agent with a free tier

Why: Built-in tools cover the usual agent needs: file operations, shell commands, web fetching and Google Search grounding.

Gained about 200 stars in the last 7 days (107.3k → 107.5k); last updated 10-19

MCP server that lets agents drive iOS and Android devices, simulators and emulators

Why: One tool set for iOS and Android across simulators, emulators and real devices.

Gained about 600 stars in the last 7 days (9.1k → 9.7k); last updated 10-18

Builds a markdown knowledge graph of your codebase so coding agents explore less

Why: Turns a repo into a graph of linked markdown notes the agent can read and grep, with no embeddings or index to keep warm.

Gained about 400 stars in the last 7 days (10.0k → 10.4k); last updated 10-18

Real-time global news and risk dashboard with maps, AI briefs and an MCP server

Why: Open to programmatic use: MCP server, REST API with OpenAPI spec, CLI and Python and Ruby SDKs.

Gained about 300 stars in the last 7 days (88.4k → 88.7k); last updated 10-18

Claude Code skills that run ML research loops with a second model as reviewer

Why: Human-in-the-loop checkpoints are configurable, from step-by-step approval to full autopilot.

Gained about 300 stars in the last 7 days (17.4k → 17.7k); last updated 10-17

Ghidra plugin and MCP server that let LLMs reverse engineer binaries

Why: Lets an LLM decompile and analyse binaries in Ghidra, which is free to use.

Gained about 200 stars in the last 7 days (11.2k → 11.4k); last updated 10-10

Chinese study guide that links AI agent learning resources to a job-search path

Why: Ties each topic to interview questions and résumé wording, with a bank of 1,000+ questions.

Gained about 200 stars in the last 7 days (10.8k → 11.0k)

Local-first AI video editor where agents edit a real multitrack timeline

Why: AI edits land on a real multitrack timeline, so results stay editable and undoable

Gained about 200 stars in the last 7 days (2.3k → 2.5k); last updated 10-18

Convert Korean HWP/HWPX, PDF and Office files to Markdown, with CLI and MCP

Why: Goes beyond extraction: compare, format-preserving patches, Markdown to HWPX, form filling and PII masking.

Gained about 200 stars in the last 7 days (2.4k → 2.6k); last updated 10-18

Free local video editor with captions, voice cloning and an MCP API

Why: No watermark, account or paywall, and everything runs locally and offline once models are downloaded.

Gained about 3.4k stars in the last 7 days (4.8k → 8.2k); last updated 10-16

20 MB database client for 70+ databases with an AI assistant and MCP server

Why: Separate MCP server reuses your saved connections, with an allowlist and read-only, read/write and full-access modes.

Gained about 1.3k stars in the last 7 days (25.8k → 27.1k); last updated 10-17

Connect Blender to any LLM for prompt-driven 3D modelling

Why: Can pull assets and models from Poly Haven, Sketchfab and AI generators such as Hyper3D Rodin and Hunyuan3D.

Gained about 900 stars in the last 7 days (30.5k → 31.4k); last updated 10-06

MCP server that lets LLMs control the Windows desktop with mouse, keyboard and shell

Why: Gives any LLM control of the Windows desktop (mouse, keyboard, windows, apps) without a vision model or fine-tuning.

Gained about 700 stars in the last 7 days (8.7k → 9.4k); last updated 10-16

Ghidra MCP server with 245 tools, read-write analysis and a live debugger

Why: Dynamic analysis: P-code emulation of single functions and a live debugger (dbgeng, gdb or lldb) over TraceRmi.

Gained about 300 stars in the last 7 days (5.4k → 5.7k); last updated 10-16

Python framework for role-based multi-agent workflows

Why: Two complementary models in one framework: autonomous Crews where flexibility helps, deterministic Flows where control matters.

Gained about 200 stars in the last 7 days (59.6k → 59.8k); last updated 10-17

Minimal 100-line software engineering agent

Why: Bash-only design works with any model, without tool-calling support.

Gained about 200 stars in the last 7 days (8.4k → 8.6k); last updated 10-12

Open-source, local AI presentation generator with API and MCP server

Why: Can create a template from an existing PPTX so new decks follow your brand design.

Gained about 200 stars in the last 7 days (11.0k → 11.2k); last updated 10-16

Adaptive Python web scraping framework with anti-bot fetchers

Why: One framework covers fast HTTP requests, dynamic browser pages and large crawls.

Gained about 1.4k stars in the last 7 days (86.7k → 88.1k); last updated 10-16

Python framework for transparent, production-ready agents

Why: Multi-agent features (pipelines, team pipelines, A2A) plus realtime voice agents.

Gained about 400 stars in the last 7 days (33.0k → 33.4k); last updated 10-15

Open course: build a Claude Code-like agent from scratch

Why: Clear, incremental structure: each lesson adds exactly one mechanism with runnable code.

Gained about 400 stars in the last 7 days (78.3k → 78.7k)

Curated list of MCP servers, organised by category

Why: Legend icons and quality badges help with a first pass.

Gained about 400 stars in the last 7 days (96.1k → 96.5k); last updated 10-15

MCP framework that gives AI agents 150+ penetration-testing tools

Why: Brings a very broad set of security tools (150+) behind a single MCP interface for AI agents.

Gained about 300 stars in the last 7 days (12.6k → 12.9k)

Curated list of OpenAI Codex plugins, skills and a repo marketplace

Why: Can be added to Codex as a repo marketplace, so entries are installable without copying files by hand.

Gained about 220 stars in the last 7 days (521 → 741); last updated 10-16

GitHub's official MCP server for repos, issues and PRs

Why: Hosted remote option with OAuth means no local setup or token management for supported clients.

Gained about 200 stars in the last 7 days (33.5k → 33.7k); last updated 10-15

Durable execution platform for reliable workflows and agents

Why: Workflows are plain code in your language rather than a DSL or visual graph.

Gained about 200 stars in the last 7 days (23.6k → 23.8k); last updated 10-16

Self-hosted AI assistant you chat with in WhatsApp, Slack, Telegram

Why: Your data stays local: state, memory and credentials live on your own machine. By default OpenClaw only performs a daily version check, and anonymous usage statistics are opt-in.

Gained about 600 stars in the last 7 days (391.5k → 392.1k); last updated 10-15

Let coding agents debug and inspect a live Chrome

Why: Brings real DevTools capabilities to agents: performance traces with insights, network analysis and source-mapped console errors.

Gained about 400 stars in the last 7 days (53.2k → 53.6k); last updated 10-15

ByteDance's open-source super-agent runtime for long tasks

Why: Batteries included: sandbox, filesystem, memory, skills, sub-agents and scheduled tasks come out of the box instead of being assembled by hand.

Gained about 300 stars in the last 7 days (83.6k → 83.9k); last updated 10-15

Agent operating system built around isolated workspaces with white-box memory

Why: Each project is a WorkSpace with its own files, memory and skills, so parallel projects do not pollute one another's context.

Gained about 300 stars in the last 7 days (4.0k → 4.3k); last updated 10-15

Desktop AI client for many LLM providers and assistants

Why: One app for many providers, including local models, with multi-model conversations side by side.

Gained about 200 stars in the last 7 days (52.5k → 52.7k); last updated 10-15

Self-hosted AI trading OS: strategies, backtests, paper and live execution

Why: Covers the whole loop from AI research and strategy code to backtests, paper trading, live execution and monitoring.

Gained about 200 stars in the last 7 days (12.6k → 12.8k); last updated 10-15

MCP server covering the full DaVinci Resolve scripting API

Why: Very broad coverage of the Resolve scripting API, with a per-method test status table in the README

Gained about 200 stars in the last 7 days (3.4k → 3.6k); last updated 10-15

Git-versioned map of your codebase written by your coding agent and served by a local MCP server; FSL-licensed, source-available

Why: Meant to survive a change of agent, person or conversation: one read of the index picks up where work left off

Gained about 200 stars in the last 7 days (1.4k → 1.6k); last updated 10-14

Microsoft's official MCP server for browser automation

Why: Works from accessibility snapshots, so no vision model is needed and actions are more deterministic than screenshot-based clicking.

Gained about 200 stars in the last 7 days (37.9k → 38.1k); last updated 10-08

Workflow automation platform for building AI agents

Why: AI features are first-class: multi-step agents, tool use, human approval steps and model switching without changing the workflow structure.

Gained about 700 stars in the last 7 days (206.7k → 207.4k); last updated 10-14

Compress tool output and logs before they reach the LLM

Why: Content-aware compression for JSON, code and prose, with reversible retrieval of the originals.

Gained about 600 stars in the last 7 days (74.8k → 75.4k); last updated 10-13

Assign issues to AI coding agents like teammates

Why: Strong oversight: replayable execution logs, per-run cost analytics and review gates before anything ships.

Gained about 600 stars in the last 7 days (52.2k → 52.8k); last updated 10-14

Up-to-date library docs for your AI coding assistant

Why: Directly reduces hallucinated APIs and outdated examples by grounding the agent in current, version-specific docs.

Gained about 200 stars in the last 7 days (62.8k → 63.0k); last updated 10-14

Open-source RAG engine with deep document parsing and citations

Why: Strong parsing of complex formats (layout analysis, OCR, table recognition), which is where many RAG pipelines lose quality.

Gained about 300 stars in the last 7 days (91.9k → 92.2k); last updated 10-14

High-performance AI gateway with one OpenAI-compatible API for 23+ providers

Why: Automatic fallbacks and load balancing across providers, models and API keys.

Gained about 200 stars in the last 7 days (8.6k → 8.8k); last updated 10-14

Code knowledge graph that cuts agent token use on codebases

Why: Single static binary with no runtime dependencies or API keys; data stays local.

Gained about 400 stars in the last 7 days (46.2k → 46.6k); last updated 10-14

Keep raw tool output out of your agent's context window

Why: Large context savings on tool-heavy sessions by keeping raw output in a sandbox.

Gained about 2.6k stars in the last 7 days (25.6k → 28.2k); last updated 10-13

Anthropic's role-based plugins for Claude Cowork

Why: Ready-made starting points for many business roles, each bundling skills, connectors, commands and sub-agents.

Gained about 2.4k stars in the last 7 days (27.1k → 29.5k); last updated 10-12

Hosted MCP gateway to thousands of ready-made Apify scrapers

Why: Hosted server means no local setup, and tools can be restricted to save context.

Gained about 1.0k stars in the last 7 days (10.0k → 11.0k); last updated 10-13

Sandboxed desktops and drivers for computer-use agents

Why: Isolated desktops, VMs and containers keep computer-use agents away from your main system.

Gained about 900 stars in the last 7 days (28.7k → 29.6k); last updated 10-13

Agent meta-harness for Claude Code and Codex swarms

Why: Multi-agent coordination and cross-session memory without changing how you use Claude Code day to day.

Gained about 500 stars in the last 7 days (74.1k → 74.6k); last updated 10-13

SDK for agent-native apps with generative UI

Why: Purpose-built front-end layer for agents: generative UI, shared state and human-in-the-loop out of the box.

Gained about 200 stars in the last 7 days (37.8k → 38.0k); last updated 10-13

Semantic code navigation and editing tools for coding agents

Why: Symbol-level tools (references, renames, moves) replace many error-prone text edits with single atomic operations.

Last updated 10-13

Type-safe Python agent framework from the Pydantic team

Why: Type safety end to end: validated outputs, typed tools and dependency injection.

Last updated 10-13

CowAgent: personal AI assistant for WeChat, Feishu and more

Why: Skills, long-term memory, a personal knowledge base and agent teams built in.

Gained about 200 stars in the last 7 days (47.2k → 47.4k); last updated 10-12

MCP plugin that lets an AI assistant drive the x64dbg debugger

Why: Mandatory bearer-token authentication, with token rotation in the config dialog

Gained about 200 stars in the last 7 days (2.3k → 2.5k); last updated 10-12

Fast Python framework for building MCP servers and clients

Why: Very little boilerplate: schemas and validation are generated from type hints and docstrings.

Last updated 10-12

Desktop cowork app that runs Claude Code, Codex and more

Why: One interface for many CLI agents, with parallel sessions and unified MCP management.

Gained about 200 stars in the last 7 days (33.3k → 33.5k)

Open-source platform for knowledge-base agents and RAG

Why: Strong support for private deployment and domestic Chinese models alongside global providers.

Last updated 10-12

Google's code-first Python toolkit for building AI agents

Why: Graph workflows, a Task API and multi-agent hierarchies cover both deterministic and autonomous patterns.

Last updated 10-12

ByteDance's multimodal agent stack and desktop GUI agent

Why: Operates any GUI through vision, including desktop apps without APIs.

Last updated 10-09

Industrial speech recognition toolkit with streaming support

Why: Full speech pipeline (ASR, VAD, punctuation, speakers, emotion) rather than transcription alone.

Last updated 10-10

Ghidra plugin and MCP server that let LLMs reverse engineer binaries

Why: Can automatically rename methods and data, and list methods, classes, imports and exports.

Gained about 500 stars in the last 7 days (10.6k → 11.1k); last updated 10-10

Builds a markdown knowledge graph of your codebase so coding agents explore less

Why: Rebuilds the structural graph against the working tree on every query, so it reflects uncommitted edits.

Gained about 300 stars in the last 7 days (9.6k → 9.9k); last updated 10-10

MCP server that lets agents drive iOS and Android devices, simulators and emulators

Why: Prefers accessibility trees, so many interactions need no vision model, and falls back to screenshots when labels are missing.

Gained about 300 stars in the last 7 days (8.7k → 9.0k); last updated 10-11

Ultra-lightweight self-hosted personal AI agent framework

Why: Long-term memory, scheduled automations and subagents built in; MIT licensed.

Gained about 200 stars in the last 7 days (48.8k → 49.0k); last updated 10-11

Claude Code skills that run ML research loops with a second model as reviewer

Why: Cross-model design uses a second model as an adversarial reviewer instead of Claude reviewing itself.

Gained about 200 stars in the last 7 days (17.1k → 17.3k); last updated 10-07

Local-first AI video editor where agents edit a real multitrack timeline

Why: Local-first: the timeline, local projects and manual editing do not need cloud services

Gained about 200 stars in the last 7 days (2.1k → 2.3k); last updated 10-07

Desktop app for working with AI agents on your own files, powered by OpenCode

Why: Permission prompts (allow once, always or deny) and debug exports make agent actions easier to audit.

Gained about 200 stars in the last 7 days (6.7k → 6.9k); last updated 10-11

MCP server for "vibe reversing" in IDA Pro

Why: Supports edits too: comments, assembly patches, C type declarations and stack variables.

Gained about 200 stars in the last 7 days (12.5k → 12.7k)

Ghidra MCP server with 245 tools, read-write analysis and a live debugger

Why: Dynamic analysis: P-code emulation of single functions and a live debugger (dbgeng, gdb or lldb) over TraceRmi.

Gained about 1.2k stars in the last 7 days (4.1k → 5.3k); last updated 10-09

20 MB database client for 70+ databases with an AI assistant and MCP server

Why: Separate MCP server reuses your saved connections, with an allowlist and read-only, read/write and full-access modes.

Gained about 1.0k stars in the last 7 days (24.6k → 25.6k); last updated 10-10

MCP server that lets LLMs control the Windows desktop with mouse, keyboard and shell

Why: State tool returns a combined snapshot of active apps and interactive elements with a screenshot.

Gained about 900 stars in the last 7 days (7.7k → 8.6k); last updated 10-10

Free local video editor with captions, voice cloning and an MCP API

Why: No watermark, account or paywall, and everything runs locally and offline once models are downloaded.

Gained about 600 stars in the last 7 days (4.1k → 4.7k); last updated 10-10

Real-time global news and risk dashboard with maps, AI briefs and an MCP server

Why: Can run fully locally with Ollama and starts with no environment variables or API keys.

Gained about 400 stars in the last 7 days (87.8k → 88.2k); last updated 10-10

Connect Blender to any LLM for prompt-driven 3D modelling

Why: Can pull assets and models from Poly Haven, Sketchfab and AI generators such as Hyper3D Rodin and Hunyuan3D.

Gained about 400 stars in the last 7 days (30.0k → 30.4k); last updated 10-06

Open-source Chrome extension for AI web automation

Why: Runs locally in your browser, so credentials and sessions are not shared with a cloud service.

Gained about 200 stars in the last 7 days (13.9k → 14.1k); last updated 10-02

Minimal 100-line software engineering agent

Why: Bash-only design works with any model, without tool-calling support.

Gained about 200 stars in the last 7 days (8.2k → 8.4k); last updated 10-06

Durable execution platform for reliable workflows and agents

Why: Strong fit for long-running AI agents and human-in-the-loop steps; already integrated by frameworks such as Pydantic AI.

Gained about 200 stars in the last 7 days (23.4k → 23.6k); last updated 10-09

Secure cloud sandboxes for running AI-generated code

Why: Isolated sandboxes keep untrusted, AI-generated code away from your infrastructure.

Gained about 200 stars in the last 7 days (14.1k → 14.3k); last updated 10-09

Control the Unity Editor from AI assistants via MCP

Why: Broad editor control: scenes, GameObjects, scripts, assets, tests, profiling and builds.

Last updated 10-08

Official Firecrawl MCP server for search, scrape and crawl

Why: Clean Markdown or schema-based JSON output that saves context.

Last updated 10-09

Build AI agents in a visual builder or TypeScript SDK

Why: Two-way sync between a no-code builder and a typed TypeScript SDK.

Last updated 10-09

Give coding agents your Figma designs to implement

Why: Structured design data gives far more accurate implementations than screenshots.

Last updated 10-03

CNCF AI agent for investigating production incidents

Why: Purpose-built for incident investigation, with deep observability integrations.

Last updated 10-08

Google's open-source terminal AI agent with a free tier

Why: Generous free tier with a personal Google account, and no API key management needed to get started.

Last updated 10-09

Self-hosted AI assistant you chat with in WhatsApp, Slack, Telegram

Why: Meets you where you already are: 20+ messaging channels, plus companion apps on macOS, iOS, Android, Windows and Linux that add voice, Canvas, camera and screen actions.

Gained about 300 stars in the last 7 days (391.2k → 391.5k); last updated 10-08

Open course: build a Claude Code-like agent from scratch

Why: Covers the full set of modern coding-agent mechanisms, from the basic loop to teams, MCP and goal loops.

Gained about 300 stars in the last 7 days (77.9k → 78.2k); last updated 09-28

ByteDance's open-source super-agent runtime for long tasks

Why: Batteries included: sandbox, filesystem, memory, skills, sub-agents and scheduled tasks come out of the box instead of being assembled by hand.

Gained about 200 stars in the last 7 days (83.3k → 83.5k); last updated 10-08

Python framework for role-based multi-agent workflows

Why: Two complementary models in one framework: autonomous Crews where flexibility helps, deterministic Flows where control matters.

Gained about 200 stars in the last 7 days (59.3k → 59.5k); last updated 10-08

Let coding agents debug and inspect a live Chrome

Why: Brings real DevTools capabilities to agents: performance traces with insights, network analysis and source-mapped console errors.

Gained about 200 stars in the last 7 days (52.9k → 53.1k); last updated 10-08

GitHub's official MCP server for repos, issues and PRs

Why: Fine-grained control through toolsets, individual tool selection and a read-only mode.

Gained about 200 stars in the last 7 days (33.3k → 33.5k); last updated 10-08

Desktop AI client for many LLM providers and assistants

Why: One app for many providers, including local models, with multi-model conversations side by side.

Gained about 200 stars in the last 7 days (52.3k → 52.5k); last updated 10-08

Open-source deep research agent that writes cited reports

Why: Produces long, cited reports from many sources, with source tracking for each claim.

Last updated 10-01

Code knowledge graph that cuts agent token use on codebases

Why: Large token savings on code exploration: one graph query replaces many search and read steps.

Gained about 400 stars in the last 7 days (45.6k → 46.0k); last updated 10-07

Compress tool output and logs before they reach the LLM

Why: Several integration modes (library, proxy, MCP server, one-command agent wrapping) cover most setups without code changes.

Gained about 400 stars in the last 7 days (74.2k → 74.6k); last updated 10-07

Python framework for transparent, production-ready agents

Why: Broad model support including Chinese providers such as DashScope, DeepSeek and Volcengine.

Gained about 300 stars in the last 7 days (32.6k → 32.9k); last updated 09-30

Workflow automation platform for building AI agents

Why: AI features are first-class: multi-step agents, tool use, human approval steps and model switching without changing the workflow structure.

Gained about 300 stars in the last 7 days (206.5k → 206.8k); last updated 10-07

Open-source RAG engine with deep document parsing and citations

Why: Strong parsing of complex formats (layout analysis, OCR, table recognition), which is where many RAG pipelines lose quality.

Gained about 200 stars in the last 7 days (91.6k → 91.8k); last updated 10-07

Microsoft's official MCP server for browser automation

Why: Works from accessibility snapshots, so no vision model is needed and actions are more deterministic than screenshot-based clicking.

Gained about 200 stars in the last 7 days (37.7k → 37.9k); last updated 10-07

Up-to-date library docs for your AI coding assistant

Why: Directly reduces hallucinated APIs and outdated examples by grounding the agent in current, version-specific docs.

Gained about 200 stars in the last 7 days (62.6k → 62.8k); last updated 10-07

Curated list of MCP servers, organised by category

Why: Very broad coverage, organised into clear categories.

Gained about 200 stars in the last 7 days (95.7k → 95.9k); last updated 09-27

Adaptive Python web scraping framework with anti-bot fetchers

Why: Adaptive selectors that survive website redesigns reduce scraper maintenance.

Gained about 1.3k stars in the last 7 days (84.7k → 86.0k); last updated 10-06

Keep raw tool output out of your agent's context window

Why: Large context savings on tool-heavy sessions by keeping raw output in a sandbox.

Gained about 1.0k stars in the last 7 days (24.5k → 25.5k); last updated 10-06

Sandboxed desktops and drivers for computer-use agents

Why: Isolated desktops, VMs and containers keep computer-use agents away from your main system.

Gained about 800 stars in the last 7 days (27.6k → 28.4k); last updated 10-06

Hosted MCP gateway to thousands of ready-made Apify scrapers

Why: Access to a very large catalogue of ready-made scrapers instead of a single tool.

Gained about 800 stars in the last 7 days (9.1k → 9.9k); last updated 10-06

Anthropic's role-based plugins for Claude Cowork

Why: Ready-made starting points for many business roles, each bundling skills, connectors, commands and sub-agents.

Gained about 500 stars in the last 7 days (25.9k → 26.4k); last updated 10-06

Agent meta-harness for Claude Code and Codex swarms

Why: Multi-agent coordination and cross-session memory without changing how you use Claude Code day to day.

Gained about 400 stars in the last 7 days (73.6k → 74.0k); last updated 10-06

Semantic code navigation and editing tools for coding agents

Why: Symbol-level tools (references, renames, moves) replace many error-prone text edits with single atomic operations.

Gained about 200 stars in the last 7 days (29.9k → 30.1k); last updated 10-06

SDK for agent-native apps with generative UI

Why: Purpose-built front-end layer for agents: generative UI, shared state and human-in-the-loop out of the box.

Gained about 200 stars in the last 7 days (37.6k → 37.8k); last updated 10-06