Open-source tools that replace what you pay for. And the ones rising on GitHub right now.

A multi-agent trading desk, running as LLM agents that debate before they call a trade.

TradingAgents recreates an entire research desk as a graph of specialised LLM agents: fundamental analysts, sentiment analysts, technical analysts and a risk manager, each arguing their case before a trader agent commits to a position. Built on LangGraph, it pulls real market data (Yahoo Finance, FinnHub, social sentiment) and outputs the same buy/hold/sell memo a human desk would produce, with the full debate attached so you can see exactly why. It's the clearest public demonstration that structured multi-agent debate beats a single model prompted to 'be a trader'.

One self-hosted chat UI for every model you already pay for separately.

LibreChat is a production-grade, self-hosted ChatGPT-style front end that talks to OpenAI, Anthropic, Google, Mistral and a dozen other providers through a single interface, with plugins, code interpreter, image generation, agents and multi-user accounts built in. Teams use it to standardise on one UI instead of juggling five separate chat subscriptions across departments, while keeping every conversation on infrastructure they control. It is one of the most actively maintained open-source LLM front ends and a genuine alternative to paying per-seat for a hosted chat product.

An open-source generative video engine released by the team behind HeyGen.

HyperFrames is HeyGen's open-source contribution to generative video: a pipeline covering the pieces that usually sit behind a paid API, frame synthesis, motion consistency, and rendering. It's aimed at teams who want to build or fine-tune their own video generation stack instead of routing every request through a metered third-party endpoint. Worth watching closely given who's behind it.

An open-source terminal built to do what Bloomberg Terminal does, without the six-figure licence.

Fincept Terminal is a community-built financial data terminal: market data, charting, screening, portfolio tools and news feeds in a single desktop application. It will never replicate the full depth of a Bloomberg Terminal's data licensing, but for independent analysts, small funds and finance students it covers a huge share of the daily workflow for the cost of a laptop instead of a $24,000/year seat licence.

Give it a topic, get a fully-produced short-form video: script, voiceover, footage and captions.

MoneyPrinterTurbo automates the entire short-form video pipeline: it writes the script with an LLM, generates or sources matching footage, adds AI voiceover and burns in captions, then exports a ready-to-post vertical video. It supports multiple LLM backends and TTS engines, so you are not locked into one vendor. For anyone running a content calendar of TikTok/Reels/Shorts, this replaces a genuinely painful manual pipeline.

Cloudflare's open-source framework for building AI agents that run on the edge, including inbox automation.

Built on Cloudflare's Agents SDK, this shows how to run a durable, stateful AI agent directly on Workers, triaging an inbox, drafting replies, and taking action, without a dedicated backend or a third-party agent-hosting fee. Because it runs on Cloudflare's edge network, it starts cold in milliseconds and scales without you provisioning anything. A strong reference implementation if you're evaluating build-vs-buy on an AI email assistant.

Zero-shot voice cloning and text-to-speech from a research lab, not a subscription API.

VoxCPM is OpenBMB's open-weight voice model: feed it a few seconds of reference audio and it clones the voice for zero-shot text-to-speech, with natural prosody rather than the flat cadence typical of older open TTS models. Running it yourself means no per-character billing and no audio ever leaving your infrastructure, which matters a lot the moment you're cloning a real person's voice.

A visual, node-based OSINT investigation tool for mapping people, domains and infrastructure.

Flowsint gives you a canvas where you drag in a lead, an email, a domain, a username, and chain OSINT modules together to pivot from one data point to the next, watching the investigation graph build in real time. It is aimed at researchers and security teams who currently stitch together the same lookups by hand across ten browser tabs, and turns that into a repeatable, visual workflow.

A curated library of reusable skills for Claude Code, from a well-known Chrome/DX engineer.

A growing collection of Claude Code skill definitions covering common engineering workflows: code review checklists, refactor patterns, testing conventions and more, packaged so you can drop them straight into a project instead of writing agent instructions from scratch. Useful as both a working toolkit and as a reference for how to structure your own skills well.

Open-source infrastructure for shipping OAuth and API integrations without maintaining 400 API clients.

Nango handles the unglamorous, failure-prone part of building integrations: OAuth flows, token refresh, rate limiting and syncing data from 400+ APIs (Salesforce, HubSpot, Slack, QuickBooks and more), through one consistent interface. It's self-hostable, so your customers' credentials and data never have to pass through a third party's cloud. If you've ever shipped 'Connect your CRM' as a feature, this is the piece that usually takes a team months to get right.

Open-source scheduling infrastructure, the code behind a Calendly-style booking page.

Cal.com is a fully self-hostable scheduling platform: booking pages, availability rules, round-robin routing for teams, and calendar sync (Google, Outlook, CalDAV), released as open-source with an optional hosted version. Because it's built as infrastructure rather than a locked product, teams embed it directly into their own app instead of iframing a third-party widget. One of the most-starred scheduling projects on GitHub for a reason.

A self-hosted customer-support inbox that unifies live chat, email and social DMs.

Chatwoot gives you one inbox for website live chat, email, WhatsApp, Instagram and Facebook messages, with team assignment, canned responses, and reporting, all running on your own infrastructure. It is the most mature open-source alternative to a hosted support suite, actively used in production by companies who do not want customer conversations sitting on a third-party server.

A modern, open-source CRM built to feel like the Notion/Linear generation of tools, not a 2010s form.

Twenty covers the core of what most teams actually use a CRM for, contacts, companies, deals, pipelines, custom fields, with a fast, keyboard-friendly interface and a real API. It's a serious attempt at an open-source CRM that doesn't feel like legacy enterprise software, and it's self-hostable if you'd rather not hand customer data to a platform vendor.

Open-source, self-hostable document signing, the transparent alternative to a closed e-signature vendor.

Documenso covers the actual job of an e-signature tool: upload a document, place signature fields, send for signing, get a legally-recognisable, verifiable audit trail back. Being open-source means the signing logic and certificate handling are auditable rather than a black box, which matters more for a signature tool than almost any other SaaS category.

An open-source, no-code automation builder with a growing library of app integrations.

Activepieces is a visual, drag-and-drop workflow builder in the same spirit as Zapier or Make, connect trigger to action across a growing catalogue of apps, but self-hostable and with pieces you can write yourself when an integration doesn't exist yet. It's a genuine option for teams who want no-code automation without a metered per-task pricing model.

Open-source notification infrastructure: one API for email, SMS, push and in-app alerts.

Novu centralises multi-channel notifications, email, SMS, push, in-app, chat, behind a single API and a visual workflow editor, instead of you wiring up five separate provider SDKs and building your own retry/templating logic. Self-hosting it means your notification logic and delivery logs stay on infrastructure you control.

Open-source, privacy-first forms and in-product surveys.

Formbricks covers standalone forms, NPS and CSAT surveys, and targeted in-app micro-surveys triggered by user behaviour, self-hostable so response data never has to leave your own database. A solid pick for teams that want proper survey logic and segmentation without sending customer feedback data to a third party.

Lightweight, open-source, privacy-first web analytics, no cookie banner required.

Plausible gives you the traffic, source and page-level metrics most teams actually check daily, without the invasive tracking or bloated script of a full marketing analytics suite. Because it doesn't use cookies or track personal data by default, sites running it often don't need a cookie consent banner at all, which is its own small conversion win.

The open-source Firebase alternative: Postgres, auth, storage and realtime, self-hostable.

Supabase wraps a real Postgres database with instant APIs, row-level-security auth, file storage and realtime subscriptions, giving small teams a production backend without writing one from scratch. It's fully open-source and self-hostable, or usable as a managed service, your choice, and it happens to be part of PURIST's own infrastructure for exactly this reason.

Run open-weight LLMs locally with a single command, no API key, no per-token bill.

Ollama packages open-weight models (Llama, Mistral, Qwen and dozens more) behind a dead-simple local API, so you can prototype or run production inference on your own hardware with zero per-token cost and no data leaving your network. For anything that does not strictly need a frontier closed model, this removes the metered API bill entirely.

A genuinely serious open-source investment research platform, not a toy Bloomberg clone.

OpenBB started as a Bloomberg Terminal clone and grew into something more useful: a full investment research platform with equities, options, crypto, macro and fixed-income data, a Python interface for building your own models, and an extension ecosystem. Hedge funds and independent researchers actually use this in production, it's the rare open-source finance tool that isn't a hobby project.

The library that lets an LLM actually drive a real browser, click, type, read the page, decide what to do next.

browser-use gives an LLM agent eyes and hands on a real Chromium browser: it reads the DOM, decides what to click or type, and loops until the task is done, booking a flight, filling a form, scraping a site that has no API. It exploded in popularity because it's the rare project that makes 'AI agent browses the web for you' actually work reliably instead of being a demo that breaks on the second click.

A lean framework for making multiple AI agents collaborate on a task with defined roles, like a real team.

CrewAI lets you define agents with a role, a goal and a backstory, an analyst, a writer, a reviewer, then wires them into a crew that hands work off to each other in sequence or in parallel. It's noticeably lighter and easier to reason about than heavier agent-graph frameworks, which is exactly why it became one of the most-adopted multi-agent libraries in under two years.

Describe an app in plain English, get a full, runnable codebase back.

gpt-engineer takes a project description and generates an entire working codebase, file structure, dependencies, implementation, then lets you iterate on it conversationally instead of file by file. One of the earliest and most influential "AI writes the whole app" projects, and still one of the cleanest implementations of the idea.

Microsoft's own multi-agent conversation framework, the research behind a lot of what 'AI agents' means today.

AutoGen models complex tasks as a conversation between multiple specialised agents (and optionally a human), handling code execution, tool use and multi-turn reasoning in one framework backed by an actual Microsoft Research team. Rare in that it combines serious academic rigor with a genuinely usable open-source library, most research code never gets this polished.

Secure, disposable cloud sandboxes for letting an AI agent actually run the code it writes.

E2B solves the part of 'AI writes and runs code' that everyone glosses over: where does that code safely execute? It spins up isolated micro-VM sandboxes in seconds, lets an agent run arbitrary code, install packages, and read the output, then destroys the environment, so a hallucinated `rm -rf` never touches anything real. Increasingly the invisible infrastructure behind AI coding assistants you've already used.

A persistent memory layer so your AI agent actually remembers the user between sessions.

Mem0 gives LLM applications long-term memory: it extracts, stores and retrieves relevant facts about a user or conversation across sessions, instead of the agent forgetting everything the moment the context window closes. A quietly essential piece of infrastructure that most teams building AI products end up needing and few build well themselves.

Vision-based browser automation that works even when a site redesign breaks every selector-based bot.

Skyvern automates browser workflows using computer vision and an LLM to understand a page visually, rather than brittle CSS/XPath selectors, so it keeps working when a site changes its layout, which is exactly when every traditional RPA bot breaks. A genuinely rare combination of vision models and browser automation, built by a team that has since productised it, with the core left open-source.