Tools

Alibaba's open-weight family. qwen3:4b (2.5GB) is quick and strong; 8b (5.2GB) if you have room. Text-only, good at writing, tidying text, and 100+ languages.

Open-source frontier-class model, strong at coding.

Z.ai's MIT-licensed open-weight MoE (~744B total, ~40B active), tuned for long-horizon agentic coding with a 1M-token context. Server-class, but Unsloth's 2-bit dynamic GGUF (UD-IQ2_M, ~239GB) fits a 256GB+ Mac Studio for a private, slow, tireless local agent.

Google's open-weight model family that runs on ordinary laptops. gemma3:1b (815MB) and 4b (3.3GB) fit 8-16GB machines; 4b and up are vision-capable (read images), with a 270m (292MB) tier for tiny tasks.

MiniMax's open-weight MoE (427B total / 26B active), 1M context, natively multimodal. Server-class (8x H200 BF16). Ships under the MiniMax Community License, not a standard permissive license.

NVIDIA's open 550B/55B hybrid Mamba-Transformer MoE for long-running, high-throughput agents, shipped with weights, data, and training recipes. vLLM day-0; NVFP4 runs on Hopper and Blackwell.

Tencent's open-source MoE (295B/21B active) built for agentic workflows, with a notably strong agentic score. Hy3-preview is a 256K-context, low-cost option for scheduled tool-using jobs.

Xiaomi's open-weight model family. MiMo-V2.5 is a capable 1M-context default (Intelligence Index ~49); MiMo-V2-Flash is a cheaper, faster 256K MoE for high-volume background work.

Apple's official open-source toolkit (BSD-3-Clause, shown at WWDC26) for exporting a curated catalog of Hugging Face models into Apple's on-device .aimodel format, plus Python authoring primitives, a Swift runtime package, and Claude Code/Codex/Gemini agent skills. Requires macOS/iOS 27.0+ and Xcode 27.0+. It is a curated, well-tested set, not a universal converter, and Apple is not accepting code PRs yet.

An AI agent that drives a real browser like a person, clicking, scrolling, logging in, and filling forms, to reach data a plain crawler cannot. MIT. Pairs an LLM with a browser harness.

Professional scraping framework (Node/TS first, with a younger Python port): rotating proxies, automatic retries, browser fingerprinting, and request-queue management to keep crawls from getting blocked. Apache-2.0.

Multi-agent orchestration framework.

Static-site generator for documentation; build fails on broken links.

Vercel's filesystem-first framework for durable backend AI agents: an agent is a directory (agent.ts for the model, instructions.md for the prompt, typed tool files, skills, channels, schedules). Tools can gate themselves behind human approval (needsApproval) and you test the agent with file-based evals (defineEval). Apache-2.0; production plumbing rides Vercel (Workflow, Sandbox, AI Gateway).

Open-source framework of prebuilt AI prompt patterns; pipe in any text and pick a pattern to summarize, extract, or analyze.

Points at a website, crawls its pages, renders JavaScript, and returns LLM-ready markdown or schema-structured JSON. Self-hostable or hosted API. License is AGPL-3.0 (SDKs MIT), so the core is strong copyleft, not permissive.

TypeScript framework from the Astro team for building headless autonomous agents that deploy to Node.js, Cloudflare Workers, GitHub Actions, GitLab CI, Daytona, and Render. Each agent runs in a configurable sandbox (virtual, local, or remote container), avoiding the cost of a full container per agent at scale. Actively developed (~993 commits) with no explicit stability guarantee on the API, so pin your version.

Self-hosted, self-improving open-source AI agent by Nous Research (MIT): persistent memory, chat gateways (WhatsApp, Telegram, Signal), six terminal backends (local, Docker, SSH, Singularity, Modal, Daytona), and a skills system that authors reusable SKILL.md skills from a source via /learn, following the agentskills.io open standard. Every saved skill becomes a slash command.

Framework for building LLM applications and chains.

Formerly MemGPT: agents that manage their own memory blocks, paging context in and out like an OS.

Leading RAG framework: index any document corpus into a vector store and retrieve the relevant chunks at query time.

React framework for production web apps.

Agentic video production system driven by a coding assistant (Claude, Cursor, Copilot). 12 pipelines, 52 tools, and 500+ agent skills spanning scripting, asset generation, editing, and final composition via FFmpeg and Remotion. Can run with zero paid API keys using Piper TTS (local narration), Archive.org/NASA/Wikimedia footage, and Pexels/Unsplash images. Paid model APIs improve quality but are optional. AGPL-3.0: fine for personal and internal use; commercial products built on top must disclose source under the same license.

AgentScope prospective memory engine: consolidates conversations into reminders and surfaces them on a schedule.

The veteran industrial-strength Python scraping framework: crawl millions of pages, extract with selectors, and export clean data. BSD-3-Clause, battle-tested for over a decade.

Research runtime (MIT, early alpha; arXiv 2605.10913, Stanford/Northeastern) that records an agent's run as a reversible, Git-like trace, so nothing touches your files until you accept it. Permissions live in the task's function signature (read-only or read-write per repo) and, on a supported OS, are enforced at the native syscall jail (macOS Seatbelt, Linux Landlock). Early alpha with changing APIs; study and pilot it in a sandbox, not production.

MineDojo research agent that learns reusable skills in Minecraft; the canonical reference for procedural memory in LLM agents (MIT).

Proposed open web standard (W3C Web Machine Learning CG, co-authored by Google and Microsoft) that lets a website expose its own features to in-browser AI agents as structured, callable tools via a navigator.modelContext API or declarative <form toolname> attributes, instead of the agent guessing where to click. In a Chrome 149 origin trial; not a ratified standard, and it only helps on sites that adopt it.

Canner's context layer for querying business data in natural language: you describe entities, relationships, and metrics once in a semantic model (their MDL), and any agent queries through that governed layer across 20-plus sources (PostgreSQL, BigQuery, Snowflake, Databricks, ClickHouse, DuckDB, and more), with a Rust engine on Apache DataFusion doing the translation. Apache-2.0 for the code (docs CC-BY-4.0, AGPL-3.0 held in reserve for future modules). Restructured May 2026: the old turnkey GenBI web app now lives on the legacy/v1 branch, and main is the context layer and SDK.

A single CLI that gives a coding agent eyes on the internet: it installs open upstream tools (yt-dlp, gh CLI, cookie-auth scrapers for Twitter/Reddit/YouTube/GitHub) and registers a SKILL.md so the agent knows when to use each. No paid API keys, which is the appeal. The catch the project is upfront about: several platforms work via your logged-in cookies, which are full credentials kept locally and carry a real account-ban risk, so use a throwaway account, never your main. Because it installs system dependencies and registers a skill, it is exactly the kind of thing to scan before running.

Terminal pair-programmer that maps your whole repo and makes changes as real git commits; model-agnostic via LiteLLM.

Anthropic's agentic coding CLI.

Anthropic's desktop app for Claude; hosts MCP servers via claude_desktop_config.json.

Autonomous coding agent inside VS Code: describe what you want and it writes, runs, and fixes the code.

Indexes a repo into a persistent knowledge graph (functions, classes, call chains, routes) across 158 languages via tree-sitter, then exposes it to any MCP client as a query API. Ships as a single static binary with zero dependencies. A preprint (arXiv:2603.27277, 31 real-world repos) reports ~10x fewer tokens and 2.1x fewer tool calls than file-by-file exploration at 83% answer quality. The savings come from feeding the agent less, not from the tool being smarter.

AI-first code editor.

Self-hostable personal AI second brain: point it at your documents and ask a model questions about your own stuff.

Open-source coding agent (VS Code/JetBrains extension + a separate OpenCode-fork CLI) built around modes: role-scoped agents with their own model, tools, and file access, plus an orchestrator that hands work between them. ~20k stars, MIT.

Private, self-hosted ChatGPT-style app: drop in images and PDFs and chat with any model, including Claude.

Desktop app for discovering and running local LLMs.

Local-first markdown knowledge base; great as an agent memory store.

Obsidian community plugin exposing a local REST API (port 27124) so external tools can search and patch your vault. By coddingtonbear.

Run open-weight LLMs locally with a simple CLI and API.

Self-hosted, ChatGPT-style browser UI for Ollama and other local models, with saved history and document upload.

Open-source terminal coding agent that talks to ~any model (75+ providers). Per-agent models, real permission controls, headless runs, and MCP tools. ~174k stars, MIT.

Autonomous coding agent that runs in a Docker sandbox with a shell, editor, and browser: hand it an issue, get a pull request.

Minimal, scriptable coding-agent CLI that behaves like a Unix tool: reads piped stdin, prints and exits, allowlists tools per run, provider-agnostic. By earendil-works (~60K stars, MIT).

Open-source, agentic social media scheduler; self-hosted alternative to Buffer and Hootsuite.

NVIDIA's security scanner for agent skills, tools, and MCP servers. Point it at a directory, file, repo URL, or zip and it checks 65 vulnerability patterns across 16 categories (prompt injection, data exfiltration, supply chain, excessive agency, MCP tool poisoning, and more) with fast static analysis plus an optional LLM pass. Emits a 0-100 risk score with LOW/MEDIUM/HIGH/CRITICAL severity and SARIF 2.1.0 for CI. Built on the Liu et al. 2026 study that scanned 42,447 skills (26.1% had a vulnerability, 5.2% likely malicious). Static analysis lowers risk, it does not certify safety.

Commercial voice-to-text dictation app.

Addy Osmani's curated set of 24 production-grade engineering skills for coding agents (Claude Code, Cursor, Gemini CLI), each a readable SKILL.md encoding a senior-engineer workflow across the dev lifecycle. The value is provenance: a small, inspectable baseline written by a credible source, the model for what a good, vettable skill looks like, not an exhaustive marketplace.

Show it one example of what you want and it learns the pattern and extracts the rest, no selectors. MIT. Note: effectively unmaintained (last release 2022), so best for simple static pages.

FSoft-AI4Code's open-source (MIT) Python CLI that generates architecture-aware docs (Mermaid architecture, data-flow, and sequence diagrams plus module prose) across nine languages, running on your machine so it works on private code. A research artifact (ACL 2026 paper) at ~1.3k stars, not a hardened product; its 'beats DeepWiki' numbers are the authors' own model-graded benchmark, and it loses to DeepWiki on C/C++. Not to be confused with Google's separate hosted 'Code Wiki'.

Memory layer that builds a vector index plus a knowledge graph over your documents; remember/recall/forget/improve API.

Open-source async crawler that turns a site into clean, LLM-ready markdown with no API key or account. Apache-2.0; uses a headless browser, so it handles JS-rendered pages.

Programmatic prompt optimization framework.

The basic file, shell, and text utilities (sort, uniq, tr, …) on every Unix box.

Temporal context-graph engine behind Zep: bi-temporal facts let you query what was true at any point in time.

Microsoft's utility that converts files and pages (PDF, Office docs, HTML, images) into clean markdown an LLM can use. MIT. Note: it converts content you already have, it is not a web crawler.

MCP server (Python/uvx) that drives Obsidian's Local REST API for indexed search and heading-level edits. By MarkusPfundstein (~3.8K stars, MIT).

MCP server for safe Obsidian vault access: AST-aware frontmatter parsing (gray-matter), whole-vault search/stats, and path filtering that excludes .obsidian and system files. By bitbonsai (~1.4K stars, MIT).

Personalization/memory layer for assistants: user, session, and agent-level memory across conversations.

OpenDataLab document parser that turns PDF, DOCX, PPTX, XLSX, and images into clean, LLM-ready markdown and JSON with strong OCR, the fix for garbage RAG inputs. Fine print: it moved off AGPLv3 to a custom 'MinerU Open Source License' (Apache-2.0 based, with added commercial and attribution conditions), so read the terms before a commercial build.

Local-first agent memory in a single SQLite file: vector + FTS5 keyword search, no cloud, no API key.

MCP server (Node) that reads and writes an Obsidian vault folder directly, no plugin needed. By StevenStavrakis (~713 stars, MIT).

Google Cloud's open convention (v0.1, Apache 2.0) for writing static agent knowledge as a bundle of markdown files with YAML frontmatter. Not a runtime or SDK: just files, version-controllable, that any agent can read.

Structured generation / guaranteed JSON outputs for LLMs.

Vectorless RAG, no embeddings, no chunking, no vector DB.

Python library (MIT) for asking a CSV, an Excel export, or a dataframe questions in plain English, including across multiple tables, and getting back numbers or charts. It answers by generating and running Python, so treat it as executing untrusted code and use its Docker sandbox (pandasai-docker) for anything you did not write. It pins Python to 3.8 through 3.11 (will not install on 3.12+), and its `ee/` enterprise directory carries its own non-MIT license.

Fast, local neural text-to-speech.

Declarative evals and red-teaming for prompts, agents, and RAG that run in CI, so a behavior regression fails a check instead of reaching users. Red-team module probes prompt injection, jailbreaks, PII leaks, and tool misuse. MIT. Note: OpenAI announced it is acquiring promptfoo (March 2026), so weigh the long-term open-source trajectory.

Packs an entire repository into one file (XML by default) with a manifest and token counts, ready for a single long-context model call.

Adaptive Python scraper that re-finds elements when a page's layout changes, with optional stealth (fingerprint spoofing, anti-bot bypass). BSD-3-Clause. The stealth features are dual-use, so use them only on sites whose terms and robots.txt allow it.

Hugging Face's minimal agent library (~1k lines). Its distinctive move is code-agents: the agent writes Python to act instead of emitting JSON tool calls, often more expressive and fewer tokens for multi-step work. Model-agnostic via LiteLLM. Run the model-written code sandboxed (E2B/Docker/Modal). Apache-2.0.

The most widely deployed database engine in the world: a single self-contained file, no server, in the public domain. It is the local data store inside countless desktop and self-hosted apps, and its built-in .dump / restore round-trip makes it the cleanest way to practice the one backup discipline that matters, restoring into a fresh database and proving the data still matches.

Infinite-canvas whiteboard SDK for React. Its viral 'make real' demo turns a hand-drawn UI sketch into working code via a vision model.

Fast, memory-efficient fine-tuning for open-weight LLMs (Llama, Mistral, Phi); Apache-2.0 core. Unsloth Studio UI is AGPL-3.0.

Fast front-end build tool and dev server.

High-throughput local LLM inference and serving engine.

Fast local speech-to-text (OpenAI Whisper in C/C++).

Alibaba Tongyi Lab's in-process vector database (Apache-2.0), the SQLite of vector search: it embeds directly into your app with no server to run. Wraps the production Proxima engine, does dense and sparse vectors, hybrid search (similarity plus structured filters), and write-ahead logging for crash safety. Python/Node/Rust/Go bindings. In-process means one writer at a time and no clustering, so reach for a server like Milvus or Qdrant when you need many writers.

Commercial social media scheduling tool.

OpenAI's consumer subscription.

OpenAI's agentic coding tool.

Remote MCP server that serves up-to-date library and framework documentation to coding agents.

Cognition's (makers of Devin) auto-generated wiki for any public GitHub repo: swap github.com for deepwiki.com to get diagrams, prose, and an 'ask' chat, free for open source, no sign-up (private repos need a Devin account). It also exposes a remote MCP server so your agent can pull repo context, with tools read_wiki_structure, read_wiki_contents, and ask_question. Only as current as its last index, and it inherits RAG's confident-stale-recall failure, so scope it to reading and orientation.

Commercial AI text-to-speech.

Free access to 45+ models (including frontier ones like GPT-5.5 and DeepSeek R1) behind your existing GitHub login, no extra signup or card. Free dev limits are small: roughly 10 requests/min and about 50 requests/day on top models (around 150/day on lighter ones), each request capped near 8k tokens in and 4k out. Sized for prototyping and model comparison, not production.

Mistral's API platform with a free Experiment tier covering the full model lineup (including the Codestral coding model and Pixtral vision model) at no cost. The shape is unusual: a very low rate ceiling (about a couple of requests per minute) paired with a large monthly token budget, so it suits patient overnight batch jobs, not live chat. Mistral no longer publishes exact free numbers, so check your account Limits page.

Google research/synthesis tool grounded in your sources.

One OpenAI-compatible API in front of hundreds of models, with a free tier (model IDs ending in :free): 20 requests/min and 50 requests/day, raised to 1000/day once you have added $10 in credits one time. Paid per-token and BYOK models too. Free models may retain or train on what you send, so keep sensitive data on a cheap paid model.

AI gateway with a YAML+CEL Routing DSL that fans a request out to a panel of models and picks or fuses the result via an arbiter (best_of_n, synthesize, majority, first, tests_pass). BYOK at provider cost. OrcaRouter-Lite is the MIT self-hosted edition (model=auto); the fan-out DSL is the hosted tier.

AI research/search subscription.

Sakana AI's orchestration 'model': one OpenAI-compatible API (Fugu, plus the deeper Fugu Ultra) that is actually a trained coordinator assembling a pool of other labs' frontier models, assigning roles, verifying, and synthesizing one answer. Built on the TRINITY + Conductor ICLR 2026 papers. The routing is a black box you cannot inspect; intelligence is borrowed from the underlying models. Not available in the EU/EEA yet.

Newsletter platform with an undocumented public archive endpoint (/api/v1/archive + per-post) that exposes post titles, dates, engagement, and full body text (no subscriber data).

Frontend cloud for deploying and previewing web apps; generous free tier.

Alibaba's Wan family of text-to-image and text-to-video models, served via API and very cheap per image or per second. Strong on general scenes and aesthetics, weaker than frontier models on text inside images and precise layout. Note: earlier releases (Wan 2.1/2.2) shipped open weights, but Wan 2.5 launched API-only, so check the specific version before assuming you can self-host.

A curl build whose TLS and HTTP/2 handshake matches a real browser's fingerprint, so requests are not rejected for not looking like a browser. MIT. Dual-use; the actively maintained successor most people use now is curl_cffi (Python). Respect each site's terms and robots.txt.

Secure cloud sandboxes for running AI-generated code: a clean Python/JS SDK (e2b_code_interpreter) that drops isolated execution into an agent in a few lines, so model-written code runs off your laptop and your prod box. Apache-2.0; hosted (free tier has limits, self-hosting the sandbox stack is non-trivial).

Fast graph database; runs in one Docker command, used by Graphiti.

Stacks the free tiers of 16 LLM providers (Google Gemini, Groq, Cerebras, Mistral, Cohere, NVIDIA, OpenRouter, GitHub Models, Cloudflare, Z.ai, HuggingFace, and more) behind one OpenAI-compatible endpoint with a prioritized fallback chain that auto-switches when a provider hits its daily cap. Self-hosted via Docker (port 3001); a single unified key fronts your encrypted provider keys. Combined free allowances reach a theoretical ~1.7B tokens/month. Self-labeled personal experimentation only: top models have the smallest daily caps, so effective quality drops late in the day and resets at midnight UTC.

Low-cost European VPS hosting, popular for self-hosting agents.