Repository navigation
FAQ
Questions about HashCortx, a free, open-source, local-first AI desktop app under the MIT license. Answers describe the current source on main; the latest release is v3.3.0.
A desktop app that puts chat, a coding agent (HashCoder), agent teams (Agent Swarm), a 3D CAD forge, an ERP builder, a finance reader and a code sandbox in one window, working with 21 cloud providers on your own keys or with models on your own machine. It is built with Tauri v2, Rust and plain JavaScript.
Seif Hashish, an independent developer and pharmacist. He is the author and maintainer.
No. It works with keys from many AI providers and is not affiliated with, sponsored by or endorsed by any of them.
The first release, v2.0.0, came out in May 2026. The latest, v3.3.0, came out on 19 October 2026. Every release is on the releases page.
Yes. There is no paid tier, no subscription and no usage cap.
Yes, if you use cloud models: you bring your own keys and pay each provider directly. With a local model you pay nothing.
MIT. You may use, modify and redistribute it, including commercially and in forks, as long as the copyright notice is kept.
There is no HashCortx server. Your code and files leave your machine only inside a request to the model you chose: with Ollama or another local model app, that stays on your machine or your network.
No analytics, no crash reports and no update checks. Every connection is for work you asked for, and SECURITY.md lists each address.
In the app's own folder on your computer, protected by your user account. They are not encrypted and not in the macOS Keychain. Why.
Yes, with Ollama or another local model app: chat, HashCoder, the knowledge base, 3D Forge and spreadsheet import all work with the network off. Web search, cloud models and the Python sandbox's first start need the internet.
That depends on the permission mode you pick. Manual, where it starts, asks before every change. Accept edits writes files in your project freely. Auto also runs a fixed list of read, search and check commands without asking. Reading a web page, and anything outside your project, asks either way.
Yes. Every decision is written to ~/.hashcortx/audit.log, and Settings has a button to view it.
A blocklist compiled into the Rust side refuses credential folders and system directories, whether a path is named directly or inside a command. On macOS, agent commands also run in the system sandbox, kept out of your keys, keychains, shell start-up files and the app's own data. The folder you opened is a boundary in Rust too: a path outside it needs your approval.
Read SECURITY.md first. The build is not code-signed, which may conflict with company policy, and keys are stored unencrypted.
macOS on Apple Silicon, built and used daily, and Windows 10, tested. Linux builds and passes CI but has not been run by anyone yet. Intel Macs and Linux build from source.
Download the DMG from the latest release and drag HashCortx to Applications. On first launch, right-click the app, choose Open, then Open again. If macOS still refuses, run:
xattr -dr com.apple.quarantine /Applications/HashCortx.app
Run the installer from the latest release. At the SmartScreen warning, choose More info, then Run anyway. The Windows build leaves out the knowledge base's search model, so it starts on any 64-bit PC and searches by keyword.
The build is not signed or notarised by Apple. That is stated in every release.
The macOS DMG is about 42 MB. Most of that is the search model for the knowledge base, which ships inside the app.
The Windows installer from the releases page does: it leaves out the search model, so it starts on any 64-bit PC. A build you make yourself includes the model by default, and that needs a processor with AVX2 (Intel Haswell, AMD Excavator or newer); on an older one, build without it, as the README shows.
Seven: Chats, HashCoder, Agent Swarm, 3D Forge, ERP, Finance and Sandbox. Agents, Split and Connections sit in the chat toolbar.
Nine: HashCortx, HashCortx Lite, Researcher, Deep Research, Coder, URL Reader, Published Papers Researcher, Medical Lexi-Check and ATS CV Auditor. You can build your own in Agents.
Yes, in HashCoder, on the project folder you open. Every edit is a diff you keep or undo, commands run in the built-in terminal, and both follow your permission mode.
It runs your tests after a change, reads a website the way a browser does before calling it done, and sends itself back for what it missed. It is fitted to the model: light mode under 5B parameters, focused tools from 5B to 15B, and the full harness for larger and cloud models.
Retrieval on your own computer. Files you attach in chat, with the knowledge base switched on in Agents, are split into passages and indexed by keyword and by meaning, using a search model built into the app. Chat's agents and HashCoder (on larger models) look things up in it, and /inject in chat adds the best passages to a message sent to a local model.
Runs a team of agents on any mix of models. Build it on a canvas, start from a template, or describe the task and have one designed. Watch every agent in the live trace, open each agent's part and the files it made, and run again the agents that did not finish.
Turns a description into real parts with real dimensions, which you can edit, fuse into one solid and export as STL, OBJ, 3MF or STEP in millimetres.
You describe your business; its agent builds a working app with records, stages and books, exported as one HTML file. You can ask it to change the app, record what happened, or answer from your records.
Reads statements, CSV, PDF and XLSX files into figures. Every total and chart is worked out by the app, not by the model.
Agents inspect untrusted code for malware, prompt injection and suspicious logic.
Yes, in a sandbox in the app, with no Python installed. Its first start downloads the runtime.
Twenty-one: Anthropic, OpenAI, Google Gemini, xAI (Grok), DeepSeek, Mistral AI, Moonshot (Kimi), Alibaba Qwen, Z.ai (GLM), Groq, Cerebras, SambaNova, NVIDIA, OpenRouter, Together AI, Fireworks AI, Hugging Face, DeepInfra, Novita, Venice and Cloudflare.
Any model Ollama runs, and models served by another local model app or server that answers in the common chat format, found on its usual port or one you name.
Only for cloud providers. Add one in Settings → API keys and press Test. A local model needs none.
Yes. Split answers one prompt with two models side by side, Agent Swarm teams can mix models, and when a model is busy or out of quota the app can move to another one on the same side, local or cloud.
Use a model that fits in your computer's memory: one that does not leaves part of itself in main memory, where every word is slower. HashCortx already sizes each request so the model stays on the graphics chip where it can, and keeps it loaded between the steps of a run. Settings → Local model gives the two Ollama settings, flash attention and a smaller form for the conversation's memory, that let more of a model fit, for a Mac and for Windows, ready to copy.
For hard coding work the large cloud models are usually stronger. HashCoder adjusts what it asks of a smaller local model, and a local model keeps everything on your machine.
Right-click it, choose Open, then Open again, or run xattr -dr com.apple.quarantine /Applications/HashCortx.app.
Press Test beside it in Settings → API keys. If it is refused, check the key and your quota on the provider's own site, which the settings link to.
Make sure Ollama is running (ollama list in a terminal), then refresh the model list. On Windows, run setx OLLAMA_ORIGINS "http://tauri.localhost" and reopen Ollama, so it accepts the app.
The processor probably lacks AVX2, which the built-in search model needs. Build without it, as the README shows, or use the installer from the releases page on Windows.
Code signing is planned. Until then, every release says how to open the app.
The Unreleased section of the changelog, Discussions and Releases.
Yes. CONTRIBUTING.md covers setup and the project's rules: no telemetry, no bundler, no framework switch, and pull requests that show they were reviewed and tested.
Bugs: open an issue. Questions and ideas: Discussions.
For anything that could put users at risk, open an issue saying only that you have a security finding and asking for a private channel, with no details. The rest is in SECURITY.md.
See also: Home · How HashCortx compares · GitHub repository