#10
The **ai-job-search** repository by Mads Lorentzen is an AI-powered job application framework that runs locally on a user's machine, built on top of Anthropic's Claude Code CLI. The project automates and structures the entire job search lifecycle — from scraping job portals and evaluating fit, to drafting tailored CVs and cover letters, preparing for interviews, and tracking application outcomes. It is notably not affiliated with Anthropic and carries no cryptocurrency or paid sponsorship component, which the author explicitly flags to ward off scams.
The framework's origin story is unusually credible: Lorentzen, a geophysicist who lost his position in late 2025, built and used this exact workflow for his own job search. After 69 tailored applications, 20 first interviews, and one signed contract, he began working as an AI engineer in June 2026 — making this a real-world validated tool rather than a theoretical exercise. He was transparent with employers about using it, and it reportedly sparked positive technical conversations rather than raising red flags.
The core workflow is organized around three primary commands: `/scrape`, which searches multiple job portals, deduplicates results, and ranks them by profile fit; `/apply <url>`, which runs a full drafter-reviewer pipeline using a second agent to critique and revise the generated CV and cover letter (output in LaTeX via lualatex and xelatex); and `/setup`, which builds a structured personal profile from uploaded documents, a pasted CV, or a guided interview. The workflow is described as language- and country-agnostic in its core logic, though the job portal integrations are currently built for the Danish market (Jobindex, Jobnet, Akademikernes Jobbank), with the pattern designed to be swapped for other local boards.
Beyond the core three commands, the framework offers ten additional commands that significantly extend its capabilities. `/interview` builds stage-specific interview prep packs from archived application materials, researches companies and interviewers, maps likely questions to STAR examples, and runs mock interviews — with a strict rule against inventing experience. `/outcome` records application results, archives submitted materials, tracks follow-up timing, and drafts channel-appropriate follow-up messages (drafts only, never auto-sends). `/rank` batch-scores scraped postings in parallel across five evaluation dimensions, flags deal-breakers and deadlines, and hands off directly to `/apply`. `/expand` enriches the user profile by scanning linked public sources such as GitHub repos, portfolio sites, Kaggle, and Google Scholar. `/upskill` analyzes skill gaps across tracked and ranked postings and produces a prioritized learning plan with web-searched resources. `/notion-sync` publishes a read-only pipeline view to a Notion database via the official Notion MCP server. `/gmail-sync` reads Gmail for application status signals and proposes batch updates for user approval before writing anything to the tracker.
Prerequisites include Claude Code (requiring a Claude Pro, Max, or Team subscription, or pay-per-token API credits), Python 3.10+, Bun for the job search CLI tools, and a LaTeX distribution. The project also supports alternative agent tools such as Codex, Gemini CLI, and others via an AGENTS.md file, and community forks adapting the full workflow already exist. Security is addressed explicitly: job postings are treated as untrusted input, the workflow follows no embedded instructions from postings, and a SECURITY.md document covers agentic defense details.
The project is trending on Trendshift, has a CI badge indicating active maintenance, includes a video walkthrough from a third-party YouTube channel, and accepts community contributions and Ko-fi support. Its combination of genuine personal proof-of-concept, practical modularity, open extensibility, and a well-documented command structure makes it a notably compelling and timely open-source tool in
AgentsAppsToolslocal-firstworkflow-automationsecurity
★ 53,944 total+7,759 this week
#12
Orca, developed by stablyai, is a cross-platform desktop application (macOS, Windows, Linux) that functions as an AI orchestrator designed for high-productivity developers — self-described as a tool for "100x builders." Its core value proposition is the ability to run multiple AI coding agents simultaneously, each isolated in its own Git worktree, so developers can fan a single prompt across several agents, compare their outputs, and merge the best result. Supported agents include Claude Code, Codex, OpenCode, Pi, Grok, Cursor, GitHub Copilot, Devin, Goose, Auggie, Muse, DeepSeek Harness, ZCode, MiMo Code, Amp, OpenClaude, Antigravity, Hermes Agent, Autohand Code, Charm, and more — essentially any CLI-based agent. A standout feature is the mobile companion app (available on iOS and Android) that lets users monitor and steer agents remotely, receiving notifications when tasks complete. Orca also embeds a high-performance terminal with WebGL rendering and infinite splits, a Chromium-based browser with a "Design Mode" that captures HTML, CSS, and screenshots directly into agent prompts, and native integrations with GitHub and Linear for in-app PR and issue management. SSH worktrees allow agents to run on remote machines with auto-reconnect and port forwarding. Developers can annotate AI-generated diffs inline and send feedback back to agents without leaving the app, and a full CLI (`orca worktree create`, `snapshot`, `click`, `fill`) enables scripting of all workflows. Additional features include quick open search across worktrees, account switching with usage tracking for Claude and Codex, rich repo previews (Markdown, PDFs, images), Computer Use for desktop UI automation, and notification/unread-state management. The project ships updates daily, signals strong active development momentum, and is licensed under MIT, making it notable for its breadth of agent support, parallel-execution model, and mobile-first monitoring approach in the rapidly growing AI coding tools space.
AgentsAppsToolscodingworkflow-automationmultimodal
★ 97,000 total+7,017 this week
#15
ECC is an MIT-licensed open-source "agent harness operating system" designed to give AI coding agents a coordinated engineering system and toolbox rather than just raw code-generation capability. Its core premise is that while an agent can already write code, it lacks a repeatable disciplined process; ECC installs that process once so it becomes part of how the agent operates. The workflow ECC enforces follows a defined loop: plan, test, implement, review, verify, remember, and improve. In practice this means the agent plans before building, verifies changes with tests, reviews its own work from a fresh context, persists memory of what matters, and converts repeated successes into reusable skills and workflows. The project's guiding philosophy is captured in the maxim "Optimize the context window. Persist everything else," reflecting a focus on efficient context usage combined with durable external memory. ECC works best with Claude Code today and provides native plugin paths for Codex and Qoder, while offering capability-limited adapters for Cursor, OpenCode, Gemini, Zed, GitHub Copilot, Antigravity, Qwen, and other harnesses—though the README repeatedly warns users to consult the platform support matrix before assuming feature parity across these seven-plus harnesses. The toolkit is substantial: it ships with 71 specialized agents, 311 skills, and 97 legacy command shims, plus hooks, rules, memory, continuous learning, "instincts," and AgentShield security scanning. The agents are specialized for planning, review, build repair, security, architecture, and domain work, while the skills cover TDD, research, security, documentation, frontend, data, machine learning, operations, and more. AgentShield provides scanning across prompts, hooks, MCP configuration, permissions, secrets, and agent files, positioning security as a first-class concern. Commands serve as convenient entry points even as ECC transitions toward a "skills-first" surface, and rules offer always-loaded standards selectable by language or project. Installation centers on a universal guided setup via npx (ecc-universal@2.2.3 setup), requiring Node.js 18 or newer, with Claude plugin setup additionally needing Git and Claude Code 2.1 or newer. The README includes a detailed Windows first-time walkthrough using PowerShell, verifying prerequisites before running the guided installer and confirming the ecc@ecc plugin is enabled via /plugin list. ECC 2.2 specifically added guided package setup for Claude Code, Codex, and Kimi Code. The project emphasizes installing only from official sources—its GitHub repository, the npm packages ecc-universal and ecc-agentshield, the official GitHub App, the plugin slug ecc@ecc, and the website ecc.tools—explicitly warning that third-party mirrors may contain malware. Regarding maturity and notability, ECC shows strong signals of being a trending, actively maintained project: it carries a "GitHub Trending Repository of the Day" badge and a global star-history rank badge, has multilingual README translations across roughly 16 languages (including Arabic, Portuguese, Chinese, Japanese, Korean, Turkish, Russian, Vietnamese, Thai, German, Spanish, Ukrainian, Polish, and Bengali), and spans multiple programming languages including Shell, TypeScript, Python, Go, Java, Perl, and Markdown. The project operates a dual model where the open-source core remains free forever under MIT, while ECC Pro is a hosted GitHub App supporting private repositories priced from $19 per seat per month; sponsorships and Pro subscriptions fund development, which the README notes enables a single maintainer to ship weekly updates across seven harnesses. It has an active community via Discord, and a roster of commercial partners and sponsors including CodeRabbit, Greptile, Moonshot AI (Kimi), Itô Markets, and SerpApi, with past and community sponsors also credited. The combination of a comprehensive agent/skill library, built-in security scanning, broad harness compatibility, continuous-learning memory systems, and the explicit context-window opt
AgentsToolsSafetycodingsecurityworkflow-automation
★ 283,422 total+6,087 this week
#22
DeepSeek Harness (`dsh`) is an open-source AI agent harness developed by DeepSeek AI, designed around an **everything-is-a-plugin** architecture powered by the Cordis framework. This modular approach, described in the academic paper *A Programming Paradigm for Spatiotemporal Composability*, enables highly composable and extensible agent workflows. The project provides a Web UI launchable via a simple `npx` command, making it accessible for rapid local or remote (SSH) deployment. It is currently in **developer preview**, iterating rapidly with expected breaking changes, signaling active early-stage development rather than production stability. Community engagement is encouraged through GitHub Discussions, a Discord server, and a plugin ecosystem discoverable via the `dsh-plugin` GitHub topic. Its notable aspects include DeepSeek AI's backing (a prominent AI lab), a clean plugin-first architecture, and low-friction onboarding via npm — all factors that make it a trending project in the AI agents space. The MIT license and open contribution model further support community adoption.
AgentsAppsMCPworkflow-automationlocal-first
★ 253,423 total+5,744 this week
#30
Paperclip is an open-source AI agent orchestration platform built on Node.js and React that functions as a full organizational operating system for teams of AI agents. Rather than managing individual agents in isolation, Paperclip treats AI agents as employees within a structured company hierarchy — complete with org charts, job titles, reporting lines, budgets, governance policies, and performance reviews. The core analogy the project uses is compelling: if OpenClaw is an employee, Paperclip is the company. This framing captures the platform's ambition to move beyond single-agent tooling into multi-agent organizational management at scale.
The workflow is straightforward in concept but powerful in execution: users define a high-level business goal (e.g., "Build the #1 AI note-taking app to $1M MRR"), assemble a team of AI agents filling roles like CEO, CTO, engineer, designer, and marketer, then approve a strategy, set budgets, and let the agents run autonomously while monitoring progress from a unified dashboard. The interface deliberately resembles a task manager, lowering the cognitive overhead of managing what is, under the hood, a complex system of org charts, delegation chains, budget controls, and governance rules.
Paperclip is built around four structural pillars. The first is an Agentic Task Manager that handles task declaration, approval gates, auditable workflows, and output verification via diffs, screenshots, and tests. The second is an Org Chart for Agents that defines roles, permissions, responsibilities, delegation hierarchies, scoped secrets, and governance rules for both human and AI participants. The third is Agent Employee Training, which includes a Skill Studio for designing and evaluating agent skills, active learning loops, quality metrics, performance reviews, and version history for skills. The fourth is an Agentic OS — the underlying infrastructure providing cross-provider runtime support, sandboxing, MCP server integration, SSO, RBAC, cost controls, data privacy, and opt-in tracing.
A standout feature is its broad agent compatibility. Paperclip works with OpenClaw, Claude Code, Codex, Cursor (including Cursor Cloud), Gemini CLI, OpenCode, Pi, Hermes and Hermes Gateway, Grok Build, and Kimi Code — essentially any agent that can receive a heartbeat signal. Custom processes, HTTP endpoints, and external adapter packages further extend the roster, making the platform highly provider-agnostic.
Key features include goal alignment (agents receive the business context behind their tasks), heartbeat-driven scheduling (agents wake for assigned work or on configured schedules), granular cost controls with budget thresholds and alerts at the company, agent, and project levels, multi-organization support from a single deployment, task threads that keep conversations and run history attached to work items, governance with configurable review and approval stages, mobile-ready monitoring, and app connections to services like GitHub, Notion, and Railway or custom MCP servers. The Skill Studio allows teams to write, test, version, and share skills org-wide, while ready-made team templates let users bootstrap entire agent organizations quickly. Experimental features include Agent Chat and connectors for Slack, Discord, Telegram, and AgentMail.
The project is MIT-licensed and actively maintained, with a cloud-hosted version available via waitlist at paperclip.ing. Its security posture is verified by Superagent. Community channels include Discord and Twitter/X. The repository's star history badge and active community presence suggest it is gaining meaningful traction. Paperclip is notable because it addresses a real and growing pain point: as developers and businesses run dozens of simultaneous AI coding and reasoning agents, coordination, cost visibility, and governance become critical bottlenecks. Paperclip positions itself as the management layer that makes autonomous AI organizations actually governable and scalable, which is a timely and differentiated value proposition in the rapidly maturing AI agent ecosystem.
AgentsToolsMCPworkflow-automationcodingsecurity
★ 104,239 total+3,767 this week
#36
AI Engineering from Scratch (rohitg00/ai-engineering-from-scratch) is a free, open source (MIT-licensed) educational curriculum designed to teach AI engineering through hands-on implementation rather than passive reading. The project's core philosophy is to have learners implement model internals, retrieval pipelines, and agent runtimes from the ground up, then test them, inspect failures, and keep both the code and evaluation results they produce. It is a large and comprehensive resource, spanning 542 lessons organized across 20 phases, with code examples offered in Python, TypeScript, Rust, and Julia. Learners can engage with the material in three distinct ways: on the companion website (aiengineeringfromscratch.com) with no setup required, through an AI tutor powered by a skill-capable coding agent, or by cloning the repository and running local code. The curriculum is structured around several learning paths tailored to different career goals, including "Model foundations" (starting with setup and tooling), "LLM systems" (starting with prompt engineering), and "Agents and delivery" (starting with the agent loop). Given the query focus on agents, the repository places significant emphasis on agent engineering, covering topics such as the agent loop, agent runtimes, tool schema design, and tools and protocols. Agent-related content is woven throughout multiple phases, including LLM engineering, tools and protocols, agent engineering, and infrastructure and production. The project features interactive labs, such as a gradient descent visualization that demonstrates how twenty scattered starting points converge toward a loss minimum on a quadratic loss surface, with adjustable learning rates and code comparing gradient descent, momentum, and Adam optimizers. A standout feature is its three project challenges, each with staged starters, reference implementations, and local graders. These include the Retrieval Evaluation Lab (Python), which builds query-by-query ranking comparisons using NDCG, precision, and recall to catch regressions in retrieval systems; the Agent Trace Debugger (TypeScript), which parses JSONL agent traces, separates overlapping child work from parent execution time, tracks token usage, and produces inspectable HTML timelines and JSON regression reports; and the Tool Call Firewall (Rust), which validates tool-call envelopes, checks caller roles and paths, and consumes single-use approvals bound to exact requests to prevent replay attacks and unauthorized writes. Each project emphasizes keeping reproducible, inspectable artifacts and is careful to state its limitations (for example, metrics describe judgments rather than establishing answer correctness, and the firewall does not provide persistent authorization or an OS sandbox). The project is notable for its breadth, its multi-language approach, and its strong emphasis on evaluation, failure inspection, and production concerns like security, secrets, and audit. Its maturity is signaled by the completed lesson content, the roadmap, extensive internationalization (the README is translated into more than 30 languages including Chinese, Japanese, Korean, Spanish, German, French, Arabic, Hindi, and many others), and sponsor support from companies like SerpApi and NitroStack. The AI tutor integration is a particularly modern feature: learners can install it via `npx skills add rohitg00/ai-engineering-from-scratch` and use it inside coding agents like Codex (via `start-learning`) and Claude Code (via `/start-learning`), with support for Agent Skills hosts and configurable scopes. This positions the project at the intersection of the growing trends around AI agents, agent evaluation and observability, tool use and protocols (such as MCP referenced by a sponsor), and agent-assisted learning. The repository is likely trending because it combines a timely, in-demand topic (AI and agent engineering) with a rigorous, practitioner-oriented pedagogy that stresses building systems from scratch, testing them, and understanding their failure modes. Its free and open-source nature, extensive multilingual accessibility, and integration with popular coding agents broaden its appeal to a global audience of engineers seeking practical AI skills. For those specifically interested in
AgentsEvalsModelscodingsecurityworkflow-automation
★ 71,111 total+3,499 this week
#42NEWPartial week
ESP32 iPhone Agent is an open-source project that enables an AI vision model to autonomously control a real iPhone using only a ~$7 ESP32-S3 microcontroller board — no computer, no jailbreak required. The board acts as both an AirPlay receiver (giving the AI "eyes" via screen mirroring over its own Wi-Fi hotspot) and a USB HID mouse/keyboard (giving it "hands" via AssistiveTouch), while an on-device iOS app serves as the "brain," decoding the mirrored stream, sending screenshots to any OpenAI-compatible vision model, and executing planner/actor commands like taps, swipes, and keystrokes. A key architectural leap from v0.1.0 is eliminating the Mac/Linux computer entirely — the agent loop now runs fully on the iPhone itself, with cellular data handling model API calls while Wi-Fi is occupied by the board's hotspot. The system features closed-loop correction (pointer re-homing, wiggle-based calibration, screen-settle detection), run replay with annotated traces, Dynamic Island status, Siri/Shortcuts integration, and exportable run data for further agent analysis. Developed with Qwen3.8-Flash on NVIDIA DGX Sparks, it supports any OpenAI-compatible endpoint including vLLM, Ollama, and cloud APIs. Limitations include low BLE bandwidth (~300–500 kbps, ≤5 fps), ASCII-only keyboard input, and portrait-only orientation. The project is early-stage but fully functional, notable for its ultra-low-cost hardware approach to autonomous mobile AI agents.
AgentsToolsAppsvisionon-devicemultimodal
★ 3,162 total+2,987 this week
#49
OpenMontage is an open-source, agentic video production system that transforms an AI coding assistant into a full-featured video production studio. Described as the first of its kind, the project allows users to describe what they want in plain language, after which an autonomous agent handles the entire pipeline: research, scripting, asset generation, editing, and final composition. The project is notable for a critical distinction it draws from typical AI video tools — it can produce genuine motion-clip video, not merely animated slideshows. The agent builds a corpus from free stock footage and open archives, retrieves actual motion clips, edits them into a timeline, and renders a finished piece, all within free or open-source workflows.
The system is designed to work with any AI coding assistant, including Claude Code, Cursor, GitHub Copilot, Windsurf, and Codex, making it broadly accessible to developers already using these tools. Users simply open the project in their assistant and describe their desired video in natural language. The agent then autonomously conducts live web research, generates AI images, writes and narrates scripts with voice direction, sources royalty-free background music, burns in word-level subtitles, and renders the final output. Before delivery, the system runs a multi-point self-review covering FFprobe validation, frame sampling, audio level analysis, delivery promise verification, and subtitle checks. Every provider selection is scored across seven dimensions with an auditable decision log, and every major creative decision is gated for user approval.
A standout feature is the Backlot Living Storyboard, a local dashboard that fills itself in automatically as the pipeline runs. Rather than relying on chat logs to understand what the agent is doing, Backlot provides a visual board where production stages light up in real time, the script appears as a screenplay page, scene cards animate while assets generate, and every provider decision and dollar spent is displayed. Asset generation pauses at a scene-by-scene contact sheet — showing takes, prompts, per-asset costs, and quality scores — so users can approve visuals before the final render. The board also includes a replay feature, allowing users to scrub through the entire production timeline after a run completes.
OpenMontage also supports starting from a reference video. Users can paste a YouTube video, Short, Reel, TikTok link, or local clip, and the agent analyzes the transcript, pacing, scenes, keyframes, and style to generate two to three differentiated production concepts, an honest tool path, cost estimates, and a sample before full production begins. This grounded approach avoids guesswork and gives users a clear picture of what the output will look like given their available tools and budget.
The repository showcases several impressive demo productions that illustrate the system's range. These include "SIGNAL FROM TOMORROW," a cinematic sci-fi trailer using Veo-generated motion clips and Remotion composition; "THE LAST BANANA," a Pixar-style animated short produced with Kling v3 via fal.ai, Google Chirp3-HD narration, and royalty-free piano music for a total cost of $1.33; "OBJECTS IN OVERDRIVE," a 3D showcase using custom Blender animation, Blender Eevee and Cycles rendering, and FFmpeg assembly; "How Salt Made History," a 100-second cinematic documentary blending real-world footage with original narration and hand-authored motion graphics; and "One Prompt Built This Complete 3D World," a 60-second fantasy world journey assembled from textured 3D assets with cinematic lighting and a planned camera path. These examples demonstrate the system's ability to span documentary, animation, product film, and 3D cinematic genres at remarkably low cost.
Technically, OpenMontage requires Python 3.10 or higher, FFmpeg, Node.js 18 or higher, and an AI coding assistant. Installation is straightforward via a single make setup command after cloning the repository. The project is licensed under AGPLv3, signaling a strong open-source commitment while restricting proprietary forks. It achieved the number one Repository of the Day on GitHub Trending, underscoring its immediate community impact. The project is actively documented with a providers guide, PR review guide, and a dedicated Agent Guide, and the team publishes every video to a YouTube channel with full prompts, pipelines, tools used, and costs so that anyone can reproduce the results. OpenMontage represents a meaningful step toward democratizing professional video production through agentic AI workflows.
AgentsAppsToolsmultimodalvideoworkflow-automation
★ 69,593 total+2,498 this week
#51
Hypit is an open-source, agentic AI video creation and cloning platform that enables AI coding agents — such as Claude Code and OpenAI Codex — to analyze a reference video and reproduce its entire production workflow, including footage, captions, B-roll, effects, and motion graphics, all described in a word-anchored markup language called SVML rather than timecode-based editing. The core promise is radical scalability: one command, one reference video, and the agent can generate up to 100 distinct variants, each preserving the viral structure while swapping hosts, languages, products, rankings, or hooks. Hypit introduces SVML (a structured video markup language) as the lingua franca between AI agents and the video rendering pipeline, making compositions editable, re-runnable, and modular rather than one-off renders. The system is deliberately model-agnostic and generation-optional: workflows can compile captions, motion graphics, and code-rendered visuals into finished videos without invoking any generative model, avoiding those service charges entirely. When generation is needed, it integrates with services like HypiHub, GPT Image 2, Seedance 2 Mini, and WhisperX for word-level caption alignment. Rendering is parallelized across 64 headless Chromium processes for speed. Installation is a single npx command that registers the /hypit skill into a compatible coding agent environment. The project targets high-volume content use cases: paid social ads cloned from Meta Ad Library, TikTok/Reels/Shorts viral remixes, TikTok Shop affiliate videos, AI UGC talking heads, podcast clips, street interview formats, and fully code-rendered visuals. Demonstrated example costs are strikingly low — complete multi-variant productions with AI-generated A-rolls, B-rolls, portraits, captions, and music running between $1.07 and $1.15 total. The tech stack is TypeScript 5.9, Node.js 22.15+, and pnpm 10.33, licensed Apache-2.0 with conditions. Hypit achieved the #1 Repository of the Day and #1 TypeScript Repository of the Day and Week on Trendshift, signaling rapid community traction. It is free to use with no per-render fees or watermarks; users pay only their chosen model and agent service providers. The project is notably early-stage but already feature-rich, with working examples, a quickstart guide, agent environment documentation, and an active Discord, Telegram, and X community, making it a compelling and timely entry at the intersection
AgentsAppsToolsvideoworkflow-automationmultimodal
★ 23,631 total+2,346 this week
#71NEWPartial week
DeepSeek Bot (DS Bot) is an open-source plugin for DeepSeek Harness (DSH) that transforms the DSH environment into a multi-bot platform where users can create, configure, and collaborate with AI bots. Key features include persistent bot memory, a "Chief" main bot that manages other bots and group chats, and group chat functionality where bots can delegate tasks to one another via @mentions. A shared Capabilities library supports MCP servers, skills, and packs that can be attached to individual bots. The plugin supports installation via npm, direct package download, or Docker for always-on NAS/server deployments, and runs on Windows, macOS, and Linux. All data is stored locally on the user's machine. Currently at v0.1 preview (early-stage), it is a community-driven, independent project — not an official DeepSeek product. It draws design inspiration from projects like OpenAI Codex, Hermes Agent, OpenCode, and Gemini CLI. Its notable appeal lies in combining local-first privacy, multi-agent orchestration, and broad deployment flexibility within a single plugin ecosystem.
AgentsMCPAppslocal-firstworkflow-automation
★ 2,514 total+1,523 this week
#76
Career-ops is an open-source AI-powered job search agent built to help candidates evaluate, filter, and apply to job listings more intelligently and efficiently. The project was created by Santiago Fernández de Valderrama Aparicio (GitHub handle @santifer) after months of sending CVs into silence with no responses. Frustrated by the asymmetry of AI being used by companies to filter candidates while job seekers had no equivalent tool, he built career-ops for himself, used it to evaluate 740 listings, applied to 68, secured 12 interviews, and received 1 offer — landing the job. He then open-sourced the tool so others could benefit from the same approach. Six months after landing that role, he left and now develops career-ops full time.
The core workflow is simple and local-first: a user pastes a job listing URL or text, and the agent evaluates it on their own machine. It checks whether the posting is still live before any effort is invested, scores the role against the user's actual CV, and recommends whether to apply or skip. If the fit is strong enough (above a 4.0/5 threshold), it drafts a tailored CV, cover letter, and answers to application questions. It also identifies relevant hiring contacts and drafts outreach notes. Critically, the system never auto-submits anything — the user always reviews and clicks Submit themselves. No emails are sent, no data is uploaded to any career-ops backend, and there is no telemetry. The user's CV travels only from their machine to whichever AI provider they have chosen.
The evaluation framework is structured across labeled sections A through H. It includes a role summary, a CV match score with per-requirement weighting and sourcing labels (whether a weight came from the job description's own wording, its structure, or an estimate — with estimates capped below top-band), level and compensation strategy, personalization guidance, interview preparation using the STAR+R method, a posting-legitimacy check (Block G) that flags whether a listing is still open or is a repost, and a Work Authorization signal that hard-blocks roles with explicit no-sponsorship language. The system is explicitly human-in-the-loop: AI evaluates and recommends, but the human decides and acts at every stage.
Beyond basic evaluation, career-ops offers ATS-compatible PDF CV generation tailored to each job description, using Space Grotesk and DM Sans typography rendered via an HTML and Playwright pipeline. Cover letters are generated with keyword mirroring, four interactive angle prompts covering why, problems, approach, and tone, and a draft-in-chat approval gate before PDF generation. A deep research mode surfaces company AI strategy, recent moves, engineering culture, and the angle a candidate's profile should take. A contact discovery mode identifies hiring managers, recruiters, or relevant team peers worth reaching out to. After a run of rejections, the system can also name the skill or experience gap that may be causing them.
The project is installed and run via a single npx command and integrates into AI CLI environments the user already has. Free and local model options are documented for budget-conscious users. On first launch, the agent conducts an onboarding chat to learn the user's CV, preferences, and constraints — no manual configuration required. The system explicitly refuses to fabricate CV content, with a code-level check that blocks PDFs containing numbers or facts not present in the user's source materials, though it acknowledges this check is not exhaustive and urges users to read every CV before sending.
In terms of maturity and traction, career-ops has achieved notable visibility. It was ranked the number-one trending repository of the day on GitHub via Trendshift. It has been covered by WIRED Greece and Business Insider. It is featured on Product Hunt and is part of Vercel's Open Source Program, with SerpApi as a named sponsor. The project maintains a public HIRED.md file and issue tracker where users who landed jobs using the tool leave public cards — a live, auditable record of real outcomes. A community manifesto at career-ops.org articulates the philosophy in six principles: apply better to fewer, signal over volume, evidence over keywords, a human decides, local-first, and dignity on both sides of the table. The manifesto has a live signatory count, with each signature recorded as an auditable commit. The project has an active Discord community and a contributor wall showing growth since April 2026. Sponsorship is explicitly ring-fenced from product influence — sponsors receive labeled visibility only and never appear in evaluations or recommendations. The README is available in 19 languages, signaling strong international reach and community investment. Overall, career-ops is a well-documented, ethically grounded, community-driven tool that sits at the intersection of AI agents, career tooling, and local-first software — making it highly relevant to trends in agentic AI, job market automation, and open-source developer tooling.
AgentsAppsToolslocal-firstworkflow-automationbrowser-use
★ 75,530 total+1,451 this week
#77
JiuwenSwarm is an open-source multi-agent collaboration framework developed by openJiuwen-ai, designed to automate complex tasks through natural language intent, swarm-based agent orchestration, and self-evolving skills. At its core, a Leader agent decomposes tasks and dynamically assembles specialized teams of sub-agents that negotiate and collaborate, operating either on a single machine or across distributed clusters. A standout feature called Swarmflow enables deterministic multi-stage Python-scripted workflows with human-in-the-loop (HITL) support, progressive tool exposure, team token budgeting, and real-time TUI run-tree monitoring. The system's Skill Self-Evolution mechanism automatically detects errors and user dissatisfaction signals, then refines Skill definitions without any model-weight retraining, while a Swarm Skills Hub allows users to search, install, remix, and publish reusable capability assets across the community. JiuwenSwarm integrates with a broad ecosystem of LLM providers — including DeepSeek, OpenAI, DashScope, SiliconFlow, OpenRouter, Huawei Cloud MaaS, and local models — and supports access via web UI, terminal (TUI), and popular messaging platforms such as Feishu, DingTalk, WeCom, WeChat, Telegram, Discord, Slack, and WhatsApp. Security is a first-class concern, with per-tool approval gates, file access whitelists, and sensitive operation interception built in. The project is currently in active beta (v0.2.4.beta3 as of August 2026), with rapid iteration covering cold-start latency reduction, multimodal Skill-Omni capabilities, and enhanced cluster collaboration. Its cross-platform desktop installer (Windows, macOS, HarmonyOS), pip package (`workswarm`), and rich documentation make it notably accessible, positioning it as a compelling trending entry in the agentic AI
AgentsToolsAppsmultimodalworkflow-automationsecurity
★ 7,458 total+1,442 this week
#81
T3 Code is an open-source "agent harness control surface" built by Ping.gg that lets developers control AI coding agents running on their local machine from a best-in-class mobile app (iOS and Android), a web app, and an Electron-based desktop app. It supports multiple AI coding agents — Claude Code, OpenAI Codex, Cursor, Grok Build, OpenCode, and Google Antigravity — requiring only that the user has those tools installed and authenticated. The project was inspired by existing tools like the Codex desktop app, Conductor, Claude Desktop, and Cursor Glass, but aims to be more performant, remote-ready, and truly open-source, with the explicit goal of giving users everything needed to fork and self-host. Installation is straightforward via a shell script, npx, or native package managers (winget, scoop, Homebrew, apt, AUR). The project is very early-stage, explicitly warns of bugs, and is mostly not accepting contributions yet. It is notable for unifying control of multiple competing AI agent ecosystems under one interface, making it highly relevant amid the current surge in agentic coding tools trending on GitHub.
AgentsAppsToolscodinglocal-firstworkflow-automation
★ 28,407 total+1,369 this week
#86
ReelMimic is a local, AI-powered 2D animation tool that analyzes a reference video's style — editing rhythm, shot lengths, transitions, framing, colors, and camera moves — then generates an entirely new video in that style from a user-provided brief. It supports seven 2D drawing engines: vector/motion graphics, hand-painted watercolor, crayon picture book, pixel art, paper cut-out stop-motion, whiteboard doodle, and anime cel. The workflow is multi-agent: up to six AI agents work in parallel on different video segments, with separate reviewer agents validating each shot using before/after screenshots. Users approve a storyboard plan before production begins and can annotate the finished video by timecode. It runs locally via Claude Code or Codex CLI, requires Node.js 22.18+, Python 3.10+, FFmpeg, and Chrome, and supports optional social posting through Upload-Post. New styles are defined in plain Markdown files. The project is early-stage, primarily tested on Windows, with watercolor and crayon engines being the slowest. It's notable for combining multi-agent AI orchestration with structured creative control for 2D animation generation — a compelling intersection of agentic AI workflows and accessible video production tooling.
AgentsAppsmultimodalvideolocal-first
★ 3,634 total+1,269 this week
#88
Herdr is a terminal runtime designed specifically for coding agents, providing persistent background sessions that survive SSH disconnections, client closures, and machine restarts. Built as a single Rust binary (no Electron), it manages multiple local and remote machines in one unified window, with a combined agent list and independent reconnects. A standout feature is real-time pane status tracking — every terminal pane is marked as working, blocked, or idle, so users instantly know when an agent is stuck and needs input. Herdr is agent-native: agents can interact with it via a CLI and socket API to spawn panes, prompt each other, and coordinate on blocking states. It supports popular coding agents like Claude Code, Codex, Cursor, OpenCode, and Grok without wrapping or replacing them — it simply owns their terminals. The tool supports tmux-style keyboard shortcuts alongside full mouse interaction, and is extensible via a plugin marketplace. Installation is straightforward across macOS, Linux, and Windows. The project appears actively maintained with Homebrew availability, GitHub releases, and enterprise contact options, signaling growing maturity and commercial interest in the AI agent tooling space.
AgentsToolsAppscodingworkflow-automationlocal-first
★ 44,678 total+1,230 this week
#89
Octop is an open-source, self-hosted AI assistant platform developed by TencentCloud, designed for households, small teams, and individual developers who want a powerful multi-user, multi-agent AI environment without sacrificing data privacy. Rather than relying on external cloud services, Octop runs entirely on the user's own machine, consolidating a web dashboard, CLI, IM channel integrations, and cron automation into a single process launched via `octop run`, with all state stored under `~/.octop/` using SQLite by default or PostgreSQL optionally. The platform's core value proposition is that every conversation, workspace, and credential remains local, while still offering a rich, collaborative agent experience.
At the heart of Octop is a multi-agent architecture built on the Octop Harness stack, which composes several focused runtimes: Octop Harness handles agent runtime including model routing, tools, skills, and conversation checkpointing; Octop Gateway bridges multiple IM platforms into a unified processing pipeline; Octop Memory provides hierarchical recall with full-text search that travels with an agent's workspace; and Octop Browser enables CDP-based headless Chromium automation for web tasks. All surfaces — Web UI, IM channels, and cron — are routed through a single in-process HarnessProcessor, making the system restart-safe and stateless at the process level.
A standout feature is the expert system, where each user can maintain multiple "experts" — specialized agents with their own workspaces, providers, channels, MBTI persona templates (16 personality types plus a quiz), and cron schedules. An expert library and expert market allow configurations to be shared across users within the same deployment. The AgentTeams feature, currently in beta, enables a coordinator agent to autonomously schedule and orchestrate multiple expert agents for multi-step tasks. Cloud collaboration allows one Octop instance to link to another, enabling cross-instance expert browsing and chat while keeping history on the expert's side.
Octop supports an impressive breadth of communication channels, including Feishu, DingTalk, QQ, WeChat, Telegram, Discord, and WeCom, as well as programmatic access via HTTP, SSE, and WebSocket APIs. The Connector ecosystem extends capabilities through OAuth and MCP gateways, with native support for Tencent suite products like Tencent Docs and Tencent Meeting. A knowledge base feature enables RAG over private documents, shareable within a deployment. Plugins can be installed and managed via the dashboard or CLI.
The ACP (Agent Client Protocol) integration is bidirectional: Octop can serve as an ACP server for external IDE tools like Zed and OpenCode, or it can delegate outbound coding tasks to runners such as OpenCode, CodeBuddy, Claude Code, and Codex. Additional surfaces include a remote desktop feature for live screen control across Linux, Windows, and macOS, a Browser AI+ mode for web automation, a Terminal AI+ mode for AI-assisted shell interaction, and native desktop clients for Windows, macOS, Linux, and FnOS NAS packages.
The technology stack is Python 3.12+, FastAPI with uvicorn, a React 18 + TypeScript + Vite + Ant Design frontend, APScheduler for scheduling, and tooling including hatchling, ruff, mypy, and pytest. The project is at version 1.0.2, MIT-licensed, and listed on Trendshift, indicating active community traction
AgentsAppsToolslocal-firstbrowser-usemultimodal
★ 9,945 total+1,217 this week
#92
ARTEMIS is an open-source AI agent framework developed by Google that enables AI assistants and automated test suites to control real Android devices using natural language instructions, effectively letting software interact with phones the way a human would. The project's core value proposition is bridging the gap between high-level AI reasoning and low-level mobile UI interaction, making it highly relevant to the surging interest in AI agents capable of operating real-world software environments. At its heart, ARTEMIS uses a multimodal observe-and-act loop that leverages element indices for UI targeting, falling back to coordinate-based and visual locating strategies for custom or non-standard interfaces — a "Dynamic-First, Coordinate-Fallback" pattern that makes it robust across diverse apps. It supports multiple frontier vision-language models including Gemini, Claude, GPT-4o, and Qwen-VL, making it model-agnostic and broadly accessible. A standout feature is its native Model Context Protocol (MCP) server, which allows AI coding IDEs such as Antigravity, Claude Code, Cursor, Windsurf, Codex, and VS Code to directly drive connected Android devices, collect Logcat diagnostics, and capture screenshots — turning any developer's phone into a live test instrument controllable from within their IDE chat. ARTEMIS ships with two execution profiles: "Flash," a fast reactive loop completing steps in roughly 3–5 seconds, and "Pro," a deeper reasoning mode suited for long-running exploratory or stability tests. The framework achieved over 99% task completion on Google Research's AndroidWorld benchmark, a suite of 100+ multi-step tasks, representing state-of-the-art performance. It exposes four usage modes: a web visual console with live screen mirroring and replay, the MCP server for IDE integration, a developer CLI for direct terminal execution, and a Python SDK for embedding into pytest or CI/CD pipelines with Pydantic-typed outputs. Setup is streamlined via one-click shell scripts that auto-install ADB, scrcpy, FFmpeg, and Python dependencies. The project is licensed under Apache 2.0, requires Python 3.12+, and has an active Discord community, signaling a growing ecosystem. Its combination of benchmark-leading performance, broad model support, MCP-native IDE integration, and Google backing makes it a notably trending repository in the AI agents space.
AgentsMCPToolsmultimodalbrowser-useworkflow-automation
★ 12,568 total+1,192 this week
#98
DSH Desktop is an open-source, community-maintained native desktop client for Windows and macOS built on top of DeepSeek Harness, a local AI agent framework. The project wraps a pinned upstream version of DeepSeek Harness — including its local Web UI, Host service, and plugin system — into a native desktop application, adding window management, system tray integration, auto-update, terminal support, and work-profile configuration. Crucially, the desktop shell itself is implemented as a DSH plugin, embodying the project's core philosophy: "everything is a plugin," including the desktop wrapper. This means the desktop capabilities can be composed, replaced, and extended using the same plugin mechanism as any other DSH plugin, powered by the Cordis plugin architecture. A built-in plugin marketplace (DSH Community Market) enables plugin discovery, installation, and management, with an open schema allowing any data source to participate. Mobile remote control via iOS and Android is built in from version 2.0.9 onward, allowing users to monitor and interact with desktop agents from their phones. Installation is zero-dependency — no Node.js or CLI required — with NSIS installers for Windows x64 and DMG for macOS Universal. The project is explicitly independent of DeepSeek AI (深度求索), with no official affiliation, endorsement, or upstream team involvement. It is actively sponsored by Alibaba Cloud (无影云电脑), UCloud (AstraFlow), 88API, and 立秋 API, signaling commercial interest and ecosystem momentum. A growing third-party plugin ecosystem — including UI skins, TUI clients, context dashboards, remote control plugins, and sidebar tools — underscores its rising community traction and notable trending potential in the AI agent tooling space.
AgentsAppsMCPlocal-firston-deviceworkflow-automation
★ 31,574 total+1,092 this week
#100
Learn Harness Engineering is a structured, project-based educational course hosted by walkinglabs on GitHub, focused on teaching developers how to build reliable environments around AI coding agents. The core thesis of the course is that model capability alone is insufficient for real-world engineering tasks — what matters is the harness: the surrounding infrastructure of instructions, tools, environment management, state tracking, verification pipelines, and lifecycle controls that make an AI agent behave predictably and produce working results. The course synthesizes cutting-edge industry research from OpenAI and Anthropic, drawing directly on published engineering blogs about harness design for long-running agents and agent-first development workflows.
The course is organized into 14 lectures and 8 hands-on projects, and is available in 15 languages including English, Simplified Chinese, Traditional Chinese, Japanese, Korean, Spanish, French, Russian, German, Arabic, Vietnamese, Uzbek, Turkish, Brazilian Portuguese, and Ukrainian — reflecting a deliberate effort to reach a global developer audience. This multilingual breadth is a notable signal of the project's ambition and community investment.
The conceptual framework at the heart of the course breaks a harness into five subsystems: instructions, tools, environment, state, and feedback. These five components are treated as universal primitives that apply across different AI coding products and workflows. A key pedagogical move in the course is applying this framework to reverse-engineer how real frontier products build their own harnesses, covered in a dedicated section called Frontier Harness Design Breakdowns. This section analyzes four production systems: Pi, which uses a minimal kernel with programmable expansion and context engineering; Claude Code, which employs a four-layer memory architecture, five-level compaction, hooks, and sub-agent isolation; Codex, which treats the repository as the source of truth and uses AGENTS.md as a directory page with worktree isolation; and DeepSeek, which is built around an "everything is a plugin" philosophy with capability seams and an event pipeline.
The most recently added content, dated August 2026, introduces Graph Engineering as a natural evolution beyond single-loop agent architectures. Lecture 14 explains why a single agent loop eventually grows into a graph, covering the four stacked layers of prompt, context, loop, and graph, and the four structural components of a graph: nodes, edges, shared state, and routing. It also addresses why in-loop checkpoints fail at scale due to Goodhart's Law, upward blindness, and conflict, and introduces a framework-agnostic six-step process for building a first graph. Project 08 operationalizes this with three progressive experiments: drawing a maker-checker loop as an explicit graph, adding parallel fan-out and fan-in nodes, and adding conditional rollback and human-approval nodes.
The July 2026 update introduced Loop Engineering, covered in Lecture 13, which transitions learners from manual prompting to automated agent loops using six primitives: automations, worktrees, skills, connectors, sub-agents, and external state. The generator/evaluator split is a central pattern here, and Project 07 walks through building goal loops, timer loops, and maker-checker loops with drop-in code templates. The course uses a compelling real-world experiment from Anthropic as motivating evidence: the same model (Claude Opus 4.5) given the same prompt produced a non-functional result without a harness after spending $9 in 20 minutes, but with a full planner-generator-evaluator harness, it spent $200 over 6 hours and produced a playable game. OpenAI reported analogous qualitative reliability improvements with Codex in well-harnessed repositories.
The repository also includes a skills directory with a harness-creator skill that can scaffold a production-grade harness — including AGENTS.md, feature lists, initialization scripts, and verification workflows — for a new project in minutes. A PDF build pipeline is included, with npm scripts to generate coursebooks locally and a GitHub Actions workflow to publish them as GitHub Releases. The project is MIT-licensed and supported by a Discord community. It references a companion repository, Awesome Harness Engineering, also maintained by walkinglabs. Overall, this project is notable for its synthesis of frontier AI engineering research into a structured, multilingual, project-driven curriculum that addresses one of the most practically important and underserved areas in the current AI development landscape: making agents reliably useful in real codebases.
AgentsToolsAppscodingworkflow-automation
★ 20,802 total+1,086 this week
#109
The awesome-dsh-plugin repository is a curated "awesome list" that catalogs community-built plugins for the DeepSeek Harness (`dsh`), DeepSeek's open-source agent harness. DeepSeek Harness is described as a runnable coding agent available in both Web and headless modes, built on a plugin-centric framework where essentially everything — models, tools, sandboxes, session storage, UI, and even the agent loop itself — is implemented as a plugin. This architecture allows plugins to extend the official coding agent, replace its core components, or assemble entirely new agent configurations. The repository collects plugins that are installable via the `dsh plugin add` command, each of which must declare a `dsh.bundle` manifest to conform to the official protocol. The list is explicitly client-agnostic, meaning plugins are included because they follow the official protocol rather than because they target any particular client.
A notable feature is that the list doubles as a plugin market index. On dsh version 0.2.1 and later, users can open Plugins → Discover → Index sources and add an index URL served from npm (via jsDelivr CDN) for fast loading, with GitHub Pages mirrors available at awesome-dsh-plugin.com and dshmarket.com. The Discover tab and the `plugin_market` agent tool search the same catalog entries. The index covers npm-published plugins with known versions, while GitHub-only entries can still be installed via accompanying commands. The README recommends dsh-market, an optional in-harness plugin market with a friendly UI offering one-click plugin installation, upgrades, and theme switching. It also highlights dsh-find-plugin, which lets the agent itself find plugins through chat. Several desktop clients are mentioned — dsh-desktop, and two differently-maintained deepseek-harness-desktop projects (one by anywhere-labs) — all of which bundle dsh-market so listed plugins are one click away.
The repository includes a prominent security warning: installing a plugin runs third-party code with the user's own permissions, giving it access to files, credentials, and the network. Tool approvals do not sandbox plugin code, and inclusion on the list is explicitly not a security review. Users are advised to inspect source code before installing and to test unfamiliar plugins in isolated environments. The listing criteria are deliberately minimal: a plugin is added if it installs with `dsh plugin add`, does what its one-line description claims, fits the right category, and is actively maintained. Every submission is verified against its own source before merging (for example, tool counts claimed in descriptions are actually counted). The maintainers emphasize that the list does not rank or judge plugin quality; listings are not permanent and are removed if repos disappear, go unmaintained, or break.
The plugins are organized into many categories, including AGI Architecture Exploration, UI Enhancements, Usage & Billing, Themes & Appearance, Models & Providers, Identity & Communication, Sessions & Messages, Memory, Tools & Capabilities, WSL & Windows Interop, Browser & Web, Vision & Multimodal, Voice & Audio, Docs & Rendering, Skills, Workflow & Automation, Git & Code Review, Notifications & Integrations, Development & Runtime, Security & Permissions, Remote & Mobile, Plugin Markets & Managers, and Just for Fun. The AGI Architecture Exploration section is especially rich, featuring experimental plugins such as a J-Space Cognition Suite, a prompt-polishing tool ported from prompt-optimizer, agentic answer reviewers that re-grade assistant turns with a separate LLM, quality-review auditors, executor/advisor workflows, evidence-checked research presets, adaptive-topology multi-agent systems, outsourcing-expert presets that force task delegation to subagents, white-box AGI memory architectures, a port of Claude Code Game Studios, chain-of-thought anchoring tools, a 636-expert roster across 22 divisions, proactive wake-up schedulers, simulated-life context injectors for roleplay characters, evidence-dri
AgentsToolsAppscodingmultimodalworkflow-automation
★ 19,382 total+953 this week
#115
The `anthropics/financial-services` GitHub repository is a collection of reference agents, skills, and data connectors built on Anthropic's Claude AI, targeting financial-services workflows across investment banking, equity research, private equity, wealth management, and fund administration. The project's core design philosophy is "one source, two deployment paths": every agent and skill can be installed as a Claude Cowork plugin for end-users or deployed headlessly via the Claude Managed Agents API behind a firm's own workflow engine, using the same system prompts and skills in both cases. The repository ships ten named end-to-end workflow agents — including a Pitch Agent, Market Researcher, Earnings Reviewer, Model Builder, GL Reconciler, KYC Screener, and others — each self-contained with bundled skills and deployable with a single install command. Underlying these agents are vertical plugin bundles (investment-banking, equity-research, private-equity, wealth-management, fund-admin, operations) that expose slash commands like `/comps`, `/dcf`, `/lbo`, and `/earnings` for granular, on-demand use. A rich set of MCP (Model Context Protocol) data connectors integrates Claude with twelve major financial data providers, including Bloomberg-rival platforms like FactSet, Morningstar, S&P Global via Kensho, Moody's, PitchBook, LSEG, and Daloopa, as well as document stores like Box and Egnyte. Partner-built plugins from LSEG and S&P Global extend coverage to bond relative value, swap curves, FX carry, options volatility, tear sheets, and earnings previews. The repo also includes admin tooling to deploy Claude inside Microsoft 365 apps — Excel, Word, PowerPoint, Outlook — routed through a firm's own cloud infrastructure on Vertex AI, Bedrock, or an internal LLM gateway. Everything is file-based (Markdown and JSON, no build step), making it straightforward to customize prompts, swap connectors, and add firm-specific context. Critically, the repository carries a strong disclaimer that all outputs are drafts for human review and do not constitute investment advice or execute any transactions. The project is notable for its breadth of real-
AgentsMCPToolsworkflow-automationmultimodal
★ 40,302 total+882 this week
#116
Rakazo is an open-source, self-hosted AI agent platform that lets users create and manage persistent AI teammates — bots with their own memory, routines, conversation history, and the ability to delegate tasks to peer bots or short-lived subagents. It positions itself as a bring-your-own-model, bring-your-own-compute alternative to hosted AI assistant products like Grok Bot, Meta Muse, OpenAI Dots, Instinct, and Hark Pro, while also competing with open-source agents like OpenClaw and Hermes Agent. Rakazo runs across web, Electron desktop, iPhone, and Android, making it unusually cross-platform for a self-hosted agent project. Key features include voice mode (with ElevenLabs, OpenAI, Cartesia, or Fish Audio), shared and isolated computer environments, browser/terminal/file/desktop access for bots, and rich app integrations via Composio, Pipedream Connect, remote MCP, and OpenAPI tool sources. Sandbox compute providers supported include Docker, E2B, Daytona, CreateOS, Box, and Railway, giving operators significant flexibility in where agent workloads run. The tech stack is modern and TypeScript-first: React 19, Vite, Tailwind CSS, Electron, Expo, Hono, oRPC, PostgreSQL, Prisma, Better Auth, and Graphile Worker. Deployment is streamlined via a one-liner Docker Compose installer or a Railway template, with detailed self-hosting, security, and backup documentation. The project is currently in beta and supports nine UI languages. Its combination of full self-hosting, persistent agent identity, multi-platform clients, and broad compute/integration support makes it a notably ambitious and trending entry in the open-source AI agent space.
AgentsAppsMCPlocal-firstbrowser-usemultimodal
★ 4,621 total+881 this week
#117
**女娲.skill (Nuwa.skill)** is an open-source Agent Skills project by Chinese indie developer 花叔 (Huashu) that "distills" the cognitive frameworks of world-famous thinkers and entrepreneurs into runnable AI perspective skills. Rather than simple role-play, it extracts five layers from public figures: communication style, mental models, decision heuristics, anti-patterns/values, and honest limitations. The distillation pipeline runs six parallel research agents across books, podcasts, interviews, social media, critic perspectives, and decision records, then applies triple-verification (cross-domain recurrence, predictive power, exclusivity) before writing results into a structured SKILL.md file. The project has already distilled 14 figures — including Steve Jobs, Elon Musk, Charlie Munger, Richard Feynman, Naval Ravikant, Nassim Taleb, Paul Graham, Andrej Karpathy, Ilya Sutskever, MrBeast, Trump, Zhang Yiming, and Zhang Xuefeng — plus one topic skill (X/Twitter mastery), all scoring A-grade (≥85) on an independent fidelity scorecard. Skills run across 50+ compatible runtimes including Claude Code, Codex, Cursor, OpenClaw, Hermes Agent, Gemini CLI, and OpenCode, installable via `npx skills add alchaincyf/nuwa-skill`. The project is notably trending as a spiritual successor to the viral colleague-skill repo, extending the concept from distilling coworkers to distilling history's greatest minds, and pairs with a companion darwin-skill for autonomous skill evolution.
AgentsToolsAppsworkflow-automationsynthetic-data
★ 34,975 total+870 this week
#119NEWPartial week
ARTEX is an AI-driven autonomous penetration testing system built with a Go backend and Next.js frontend, designed to operate within isolated lab environments. It employs a multi-agent LLM architecture consisting of a planner agent and multiple worker agents that collaboratively explore targets using a dual-graph design: a globally shared Asset Graph (tracking domains, subdomains, IPs, services, endpoints) and a per-task Exploration Graph (tracking goals, intents, facts, findings, hints), linked via anchor points to enable asset coverage tracking. The planner reads the exploration graph's state and dispatches intents only when uncovered directions exist; workers claim one intent, execute real tools (Bash, HTTP, Kali-based), and write results back to both graphs, triggering the next planner cycle in an event-driven closed loop. A notable feature is inter-worker process-level information exchange: workers can search other workers' execution traces within the same task, enabling knowledge reuse even before facts are formally recorded. The system includes human-in-the-loop approval/interception, full traffic recording via a MITM proxy with CA verification, MCP (Model Context Protocol) support for both Streamable HTTP and legacy SSE transports, and integration with ScopeSentry for asset synchronization. It ships as a single embedded binary or Docker Compose stack, supports one-click in-UI updates with SHA256 verification, rollback, and idempotent schema migration on startup. A dedicated retester agent handles manual vulnerability re-verification. The project is at an early but functional release stage (v0.3.3), licensed under AGPL-3.0, and is notable for its sophisticated agentic architecture, dual-graph reasoning model, and production-oriented operational features like safe self-update and reverse proxy guidance.
AgentsSafetyMCPsecurityworkflow-automationbrowser-use
★ 1,265 total+855 this week
#130NEWPartial week
nanoMuse is an open-source personal agent designed to run across every device a user owns — phone, desktop, and browser — inspired by Meta's Muse but built to be self-hosted, privacy-respecting, and free. Rather than answering questions, it actively performs tasks: executing shell commands, controlling a browser, connecting to MCP servers, and even operating apps through screen interaction (called "Hands") on Android and desktop platforms where system APIs permit. The agent is persistent, continuing to work while the app is closed, running scheduled goals and morning feeds, and it always pauses to request user confirmation before irreversible actions like deletions, payments, or sends. A key architectural feature is its multi-device relay system: each device runs its own agent instance, and they communicate through a shared relay hub, so a task spoken on a phone can execute on a PC, with approvals routed back to wherever the user is. Memory is stored as human-readable Markdown files, and the agent supports a customizable visual identity — including animated avatars generated by image and video models. It integrates with chat platforms including Telegram, Feishu, DingTalk, and WeCom. Model flexibility is broad, supporting 21 providers including OpenAI, Gemini, DeepSeek, Claude, Kimi, OpenRouter, and Bailian, plus plans like ChatGPT. The project is built on OpenMinis and uses Alpine Linux under proot on Android. It reached 1,000 GitHub stars within 18 days of launch, ranked on GitHub Trending's Swift list, and its accompanying academic paper appeared on arXiv and ranked third on Hugging Face Daily Papers. Currently at version 1.0.5 ("Lookout"), it is licensed under GPL-3.0-or-later and positioned as a non-profit, community-driven alternative to closed personal AI agents. Its rapid traction, cross-platform scope, self-hosting capability, and academic backing make it a notably trending project in the agent-skills space.
AgentsToolsMCPon-devicebrowser-useworkflow-automation
★ 1,490 total+402 this week