NETROOM
Excalidraw is a virtual whiteboard for diagrams in a hand-drawn style: architecture sketches, user flows, wireframes and anything else that is quicker to draw than to describe. It runs in the browser with no sign-up - open a tab and start drawing. There are libraries of ready-made shapes, export to PNG and SVG, collaboration over a link with end-to-end encrypted rooms, and files you can keep locally or in your own storage. Open source, with a self-hosted option for teams.
Squoosh compresses images in the browser: drop a file in, pick a format and a quality, download the result. A slider puts the original and the compressed version side by side, so you can see exactly where artefacts start and stop just before them. It handles a range of formats, each with its own controls - from a single quality dial to the codec's own parameters. Nothing is uploaded: the work happens on your device, and it runs offline once installed as an app. Built by Google Chrome Labs, open source, no sign-up.
The Cyber Swiss Army Knife, built by the UK's GCHQ. You chain operations into a "recipe" and push text or a file through it: base64, XOR, AES, DES, Blowfish, hashes, hexdumps, X.509 and IPv6 parsing, archive extraction, charset conversion. Everything runs in the browser - nothing is sent to a server, and you can download the page and use it offline. Apache 2.0 plus Crown Copyright. Invaluable for picking apart logs, dumps and suspicious payloads.
A Swiss army knife for PDFs in the browser, running on your own server: merge and split, rotate, compress, convert to and from office formats, OCR, sign, redact, and chain operations through a REST API - over fifty tools in all. The point is that files never reach someone else's cloud. One docker run gets you going, and a desktop client and Kubernetes charts exist too. It is open core: the base is open source while some capabilities sit behind paid Server and Enterprise plans.
Vercel's generative UI agent: from a prompt, screenshot or mockup it produces React components with Tailwind and shadcn/ui, shows a live preview and deploys in a click. It's strongest on its home stack - Next.js App Router and server components; for Vue or hand-written CSS there's little point. Credit-based billing: $5/month free, Premium from $20, and the API only on paid plans. The output is genuinely usable but needs review - accessibility and state bugs show up regularly.
A research agent from Princeton and Stanford: it takes a GitHub issue, navigates the repo through a purpose-built text interface, edits code and produces a patch. It also handles offensive-security tasks. Any model plugs in via litellm - your key, your bill. One thing to know: the team's focus shifted to mini-swe-agent, roughly a hundred lines that scores higher on SWE-bench verified with almost no config. Start experiments there; take full SWE-agent for the heavier harness.
Replit's cloud agent: from a text description it scaffolds a project, installs dependencies, provisions a database, writes auth and deploys - all in the browser with no local setup. Agent 3 runs autonomously for hours, browser-tests its own output and spawns sub-agents. The catch is effort-based billing: the agent decides what a checkpoint costs and you learn the number afterwards. Complaints about surprise bills are common. Entry plan is $20/mo, with credits burned on top.
The original AutoGPT, except it stopped being a runaway prompt loop long ago: the project became a platform with a visual block builder, a scheduler and agents that run on a cron. The repo is alive with daily commits. One detail matters - the licensing is split: the platform is under Polyform Shield, so you cannot build a competing hosted service on it, while the rest stays MIT. You deploy it with docker compose rather than installing a package, and it does not embed as a library.
An autonomous dev agent, formerly OpenDevin: it boots into a Docker sandbox with a filesystem, terminal and browser, and carries a task from issue to pull request. One of the strongest open solutions on SWE-bench. Cloud and self-hosted options both exist, and you bring your own model - Claude usually gives the best results, but the token bill is yours and it adds up on long runs. Overkill for a one-file edit; it earns its keep on long unattended tasks.
An open-source personal agent runtime in TypeScript: it lives on your machine or server, answers in Slack, Discord, Telegram and iMessage, and handles subagents, schedules and persistent memory. Any model works - a Claude subscription, an API key or something local. It has been renamed twice (Clawdbot, then Moltbot), so guides still float around under the old names. The agent gets real system access, so sandboxing is entirely on you. Growing explosively, releasing constantly. MIT.
A local code interpreter: describe the task in plain language and the model writes Python, JavaScript or shell and runs it on your machine. Unlike ChatGPT's sandbox the file and network access is real, so run it in a container or with per-step confirmation. Any model works through LiteLLM, local ones included. The project has passed 60k stars, but development has slowed considerably - treat it as a finished tool rather than a foundation to build a product on.
A cloud-hosted autonomous agent: hand it a task and it spins up a virtual computer with a browser, terminal and filesystem, then works for hours unattended - research, data gathering, form filling, reports and decks. Unlike a chat assistant, it hands back artifacts rather than messages. It is a closed subscription service that burns credits per task, with no source and no self-hosting. Long-running jobs cost noticeably more than the pricing page suggests.
A hosted app builder driven by prompts: describe what you want and get a working React plus Supabase full-stack app, with live preview, chat-based edits and one-click deploy. The code syncs to GitHub and you can take it with you, so there is no source-level lock-in. Its strength is time to first prototype; its weakness is everything after that - complex logic and performance still need a human. Credit-metered subscription, with only a token free tier.
A terminal agent from Nous Research built around memory across runs: it saves what worked as reusable skills, has a cron-style scheduler and can delegate to subagents. It hooks into messengers - Telegram, Discord, Slack - and runs headless on a server, not just inside an interactive session. Bring your own model: Nous Portal, OpenRouter, OpenAI and others, switched in config. Installed by script; needs Python 3.11 and Node. MIT, and young enough that the API and skill format still move.
A personal agent that lives in the terminal and stays there: it runs shell and Python, patches files, drives a browser and reads screenshots. Install with pipx, configure with a single file - no server, no web UI. Any model works: Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter, or llama.cpp locally, so you can run with no cloud at all. It shines on a remote box over ssh and inside CI, where GUI agents are useless. Python 3.10+, MIT, on 0.32 and updated often.
Archived. The 2023 project that generated an entire codebase from a single prompt - it made noise, pulled tens of thousands of stars and set the template for 'the agent writes the app'. The repo went read-only in April 2026 and the authors point elsewhere themselves: Lovable (formerly gptengineer.app) for a managed service, aider for a CLI that is still maintained. Nothing here to put into production - the dependencies have rotted and models have turned over twice since. The value is historical now.
Block's autonomous agent: a terminal tool plus a desktop app, written in Rust, installed as a single binary. It isn't code-only - it edits files, runs commands, drives a browser, and extends through MCP with dozens of servers available out of the box. You pick the provider: Anthropic, OpenAI, Gemini, Bedrock, OpenRouter or a local Ollama, using your own key. What sets it apart from Claude Code or Cursor is that nothing is vendor-locked and it's all open, Apache-2.0. Development is brisk and the extension surface still shifts.
Cognition's cloud software engineer: hand it a task and it spins up its own VM with a terminal, browser and editor, keeps a branch and opens a PR. It shines on grunt work - dependency upgrades, repetitive repo-wide edits, spelunking through legacy code. Billing is per ACU, meaning compute time, and a hard task runs up the meter fast. Everything is closed: no local run, and your code goes to the vendor's cloud. You still review every diff by hand.
An open-source assistant for VS Code and JetBrains: autocomplete, codebase chat and custom agents defined in config and pulled from a hub. Any model works - Claude, GPT or a local Ollama - with your own key and no vendor lock. The reason to pick it is control over where your code goes: you can wire up a fully offline model stack. Development is steady, if less frenetic than Cline's. Apache-2.0, with a separate CLI alongside the extensions.
An autonomous coding agent: a VS Code extension plus a CLI you install straight from npm. It plans an edit, shows the diff before applying it, runs terminal commands and drives a browser. Bring your own key - Anthropic, OpenAI or local models; there is no subscription, you pay the provider per token. Releases land almost daily and the source is open. Choose it when you want a checkpoint at every agent step and a transparent bill. If you would rather hand over the wheel, look at Claude Code.
StackBlitz's in-browser agent: it writes the code, installs dependencies and boots a dev server inside WebContainers, with no local toolchain at all. Genuinely good for an evening prototype or a demo. The open-source repo was archived back in late 2024 - only the commercial product moves forward, and the community carries on in the bolt.diy fork. Billing is token-based and gets expensive on a real project, so export the code into your own git early.
A pair programmer that lives in your terminal: it reads the repo, edits files and commits for you, so every step is undoable with plain git. Its repo map feeds the model context without you listing files by hand. Bring your own key - Claude, GPT or local models via LiteLLM. Momentum has clearly slowed: the last aider-chat release landed in February 2026 and commits trail off in May. Still usable, but compare it against Claude Code and Cline for a long-term bet. Install with pip install aider-chat.
The TypeScript SDK for the Model Context Protocol: server and client, stdio, streamable HTTP and SSE transports, plus ready-made OAuth for remote servers. This is what MCP servers for Claude Code, Cursor and VS Code are written with, and tool schemas are declared in zod. The annoyance: express, hono and jose sit in its dependencies, which is dead weight for a tiny stdio server. Nominally 1.x, yet breaking changes land in minors - pin it. Node 18+, MIT.
The Pythonic way to write MCP servers: decorate a function, let annotations produce the tool schema, and leave transport and handshake to the library. Version 1.0 was donated to the official mcp SDK; this package builds on top of it and has since gone much further - auth, server proxying and mounting, generating a server from an OpenAPI spec, an in-memory test client, deployment. It moves fast, with minors landing almost weekly. Python 3.10+, Apache-2.0.
The official Python SDK for the Model Context Protocol, covering both server and client. Tools, resources and prompts are registered with FastMCP-style decorators; transports are stdio, streamable HTTP and the fading SSE. This is what you write integrations with so that Claude Code, Cursor, VS Code and other hosts can pick them up. Marked Production/Stable, but the 2.x major moved the API - 1.x-era tutorials no longer compile, so read the changelog. Python 3.10+, MIT.
JS/TS client for fal.ai, a hosted runner for image, video and audio models (FLUX, Veo, Kling and friends). Everything goes through a queue: subscribe waits and streams progress, queue hands you an id for a webhook. It also uploads files to their storage. The key is paid, billed per GPU second. Important: never ship the key to the browser - the package includes a Next.js proxy route. Node 18+, MIT. It supersedes @fal-ai/serverless-client, which no longer gets updates.
A Python client for Stability AI's gRPC API - Stable Diffusion generation and upscaling. Don't pick it up: the last release was May 2024, the repo is quiet, and Stability moved its platform to the REST v2beta API, which this SDK doesn't speak. For new integrations just call api.stability.ai with an ordinary HTTP client or take a current wrapper. Keep the package only if it's already in your project and still talking to the old endpoint.
A Rust, GPU-rendered terminal that grew into an agentic dev environment: blocks instead of an endless log, real command editing, and on top an agent that runs multi-step CLI tasks with per-step approval. Project rules live in WARP.md, and cloud background agents handle long jobs. The source is now open under MIT/AGPL, but the service remains: 75 free credits a month, Build at $20. Paid plans support bring-your-own OpenAI or Anthropic keys so inference is billed by the provider.
Editors' picks