Repository navigation
Home
HashCortx is a free, open-source, local-first AI desktop app. It puts chat, a coding agent, agent teams, a 3D CAD forge and an ERP builder in one window, on your own keys and your own machine, with no backend, no telemetry and no account.
- Repository: Hash-7777/HashCortX
- License: MIT
- Latest release: v3.3.0, the focus release (19 October 2026)
- Runs on: 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 size: the macOS DMG is about 42 MB, most of it the search model for the local knowledge base.
- Built with: Tauri v2, Rust and plain JavaScript, with no bundler and no minified app code.
- Author: Seif Hashish
This wiki describes the current source on main. Anything newer than v3.3.0 is listed under Unreleased in the changelog and arrives with the next release.
| Workspace | What it does |
|---|---|
| Chats | Every provider in one conversation, with projects, attachments, memory, a local knowledge base, a queue for messages sent while a model works, and versions of every edited message and regenerated answer |
| HashCoder | A coding agent on your real project: file tree, diffs you keep or undo, a terminal, and a harness that checks its work |
| Agent Swarm | Teams of agents on any mix of models, built on a canvas, started from a template or designed from your task |
| 3D Forge | Describe a part and get a dimensioned solid, in colour, exported as STL, OBJ, 3MF or STEP in millimetres |
| ERP | Describe your business and its agent builds a working app with records, stages and books, exported as one HTML file |
| Finance | Statements, CSV, PDF and XLSX read into figures, every total and chart worked out by the app |
| Sandbox | Agents that inspect untrusted code for malware, prompt injection and suspicious logic |
Beside them in the chat toolbar: Agents, nine built-in specialists and a builder for your own; Split, one prompt answered by two models side by side; and Connections, business systems connected over MCP (Odoo, GitHub, Stripe and Supabase are ready-made), whose keys never reach a page or a model and whose every change asks you first.
The nine built-in agents: HashCortx, HashCortx Lite, Researcher, Deep Research, Coder, URL Reader, Published Papers Researcher, Medical Lexi-Check and ATS CV Auditor.
Twenty-one cloud providers, with your own keys: 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.
On your own machine: Ollama, and any other local model app or server that answers in the common chat format, found on its usual port or one you name. With a local model the whole app works with the network off, except the Python sandbox's first start.
A local model is kept on your graphics chip where it fits: the app sizes each request's window to what keeps it there, measured after every answer, and keeps the model loaded between the steps of a HashCoder run. Smaller models read one line per tool, and a model without tool calling of its own answers in one shape that Ollama holds it to. Settings → Local model gives the two Ollama settings that let more of a model fit, for a Mac and for Windows.
HashCoder works on your real files. Every edit is a diff with Keep and Undo, tests run in the built-in terminal, and screenshots, PDFs and files attach by paste or drop. What the model is shown, what it may do without asking and what it has to prove before a run may end are decided by the app, not by the model.
- It proves its work. It runs your tests after a change, reads a website the way a browser does, and sends itself back for what it left unfinished.
- It fits the model. Under 5B parameters, light mode, where the model writes whole files and the app does the tool work; 5B to 15B, focused tools with the steps written out; 15B and up and cloud models, the full harness with a map of your code and a reviewer for larger changes.
- It keeps what you did not ask to change, and it does only what the request asks.
- It asks as much as you choose. Manual asks before every change; Accept edits writes project files freely; Auto adds a fixed list of read, search and check commands.
HashCortx answers from your own files without sending them anywhere to be indexed.
- With the knowledge base switched on (in Agents), a PDF or text file attached in chat is split into overlapping passages and stored on your computer.
- Every passage is indexed by its keywords and as a meaning vector from
bge-small-en-v1.5, a search model built into the app and run in its Rust core. Nothing is downloaded and nothing is uploaded to index or search. - A question is ranked against both and the two rankings are fused, so a search finds passages that mean the same thing as well as exact names and codes.
- Chat's agents look things up with a
search_knowledgetool, and so does HashCoder on larger local and cloud models. In chat,/injectadds the best passages to a message sent to a local model. - The Windows build leaves the search model out and searches by keyword only.
- Nothing phones home. No analytics, no crash reports, no update checks, and no HashCortx server: requests go from your computer to the service you chose.
-
You approve what matters. Commands, deletions, web pages and anything outside your project wait for your yes, and every decision is written to
~/.hashcortx/audit.log(Settings has a button to view it). - A blocklist compiled into Rust refuses keys, credentials and system folders, as files or inside commands.
- On macOS, agent commands run in the system sandbox, kept out of your keys, keychains, shell start-up files and the app's own data.
- A task on a local model stays local, and cloud requests are capped at 30 a minute and 6 at once.
- What it is not: your API keys are stored unencrypted in the app's own folder, protected by your user account, not by the macOS Keychain; and the build is not code-signed.
The full account, including every address the app connects to, is in SECURITY.md.
macOS (Apple Silicon). Download the DMG from the latest release and drag HashCortx to Applications. The build is not notarised: 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 in Terminal.
Windows. Run the installer from the latest release. Windows shows a SmartScreen warning the first time: choose More info, then Run anyway. To use Ollama, run setx OLLAMA_ORIGINS "http://tauri.localhost" and reopen Ollama.
Linux and Intel Macs. Build from source, as the README describes.
Then open Settings → API keys, add a key and press Test, or run a model on your own machine with Settings → Local model.