-
Notifications
You must be signed in to change notification settings - Fork 8
FAQ
Frequently asked questions about HashCortx — a free, open-source, local-first AI desktop app for developers, available under the MIT license.
HashCortx is a native macOS AI desktop application that bundles 11 specialized AI workspaces — a coding agent, multi-provider chat, multi-agent swarms, 9 pre-built specialist agents, financial document analysis, security scanning, 3D planning, and more — into a single 8.9 MB app built on Tauri v2.
HashCortx is built by Seif Hashish, an independent open-source developer with a pharma and clinical background. He is the sole author and maintainer.
No. HashCortx is an independent open-source project. It works with API keys from many AI providers but is not affiliated with, sponsored by, or endorsed by any of them.
HashCortx v2.0.0 was released in May 2026.
Yes. HashCortx was developed with heavy AI assistance — approximately 30 million tokens consumed across Claude, GPT, and other frontier models during the build of v2.0.0. Architecture, product direction, mode design, the security model, and final code review were directed by the human author; AI handled the bulk of implementation, refactoring, and iteration under that direction.
Transparency. HashCortx is an AI tool itself — using AI to build it and hiding that fact would be inconsistent. The token count is shared so users can judge the codebase honestly. All source is open at Hash-7777/HashCortX and reviewable line by line.
HashCortx ships after extensive testing and human review. Bugs exist in all software — AI-assisted or hand-written — and can be reported via GitHub Issues. The Permission Guard, Audit Log, and source-grounded modes are deliberate safeguards designed to make AI-generated output verifiable at runtime — the same scrutiny applies to the app itself.
Yes. Free to download, free to use, free to modify, free to redistribute.
No. There is no paid tier, no subscription, no usage cap, and no premium features. Everything is included.
Yes if you use cloud providers — you bring your own API keys and pay each provider directly. You can avoid all AI costs by using local models via Ollama.
MIT. Use, modify, and redistribute freely, including in commercial products, as long as you preserve the copyright notice.
Yes. MIT permits commercial use.
Yes. MIT permits forking and redistribution.
No. HashCortx has no backend server. AI requests go directly from your machine to the provider you configured. Nothing passes through HashCortx infrastructure — there is no HashCortx infrastructure.
No analytics, no tracking, no usage reporting, no error-reporting backend. Verify in the source.
In the macOS Keychain. Never on disk in plaintext, never in config files, never transmitted anywhere except the corresponding provider's API.
Yes, using local models via Ollama. Cloud providers require internet access.
HashCortx makes no network calls except to the AI provider endpoints you configure. No backend, no auto-update, no analytics. The v2.0.0 build is unsigned, which may conflict with corporate Gatekeeper policies — check with your security team and review the open source code first.
Filesystem and shell calls from the coding agent are intercepted by a denylist-based gatekeeper before execution. Each new request prompts the user; every guarded action is recorded in the built-in Audit Log.
A persistent log of every guarded action — file reads, file writes, shell commands — performed by the coding agent. Viewable in Settings → Audit.
macOS on Apple Silicon (M1, M2, M3, M4) in v2.0.0. Intel Mac, Windows, and Linux builds are planned but not yet released.
Not in the shipped build. The codebase compiles for Intel but no Intel binary is distributed in v2.0.0.
Not yet. Tauri supports both platforms; release builds for them are planned.
Download HashCortx-2.0.0-macOS-arm64.dmg from the latest release, open the DMG, drag the app to /Applications, and on first launch right-click → Open → Open.
The v2.0.0 build is unsigned. Right-click the app → Open → Open. Or run:
xattr -dr com.apple.quarantine /Applications/HashCortx.app
Code signing is planned.
8.9 MB. Roughly 30 times smaller than typical Electron-based AI desktop apps (100–300 MB).
Idle usage is typically under 100 MB. Code mode with a large project file tree and long chat history can push it higher.
Chats, Agents, Code (HashCoder), Split, 3D Forge, Finance AI, Sandbox, ERP / Systems Builder, Agent Swarm, Virtual OS, Agent Maker.
Nine: Personal Assistant, Quick Assistant, Research Agent, Deep Research, Senior Engineer, Page Analyzer, PubMed Agent, Drug Interaction, ATS CV Auditor.
Yes. Agent Maker lets you build custom agents with name, icon, system prompt, and curated tools — no code.
Yes, in Code mode. It uses real file tools (readFile, writeFile, listDirectory, runShell) gated by the Permission Guard.
Yes, in Code mode, gated by the Permission Guard with the same approve-and-audit flow as file operations.
Yes. Code mode includes a browser panel. The Page Analyzer agent fetches and reads URLs you provide.
Yes. Finance AI mode analyzes bank statements, CSVs, PDFs, and XLSX files. Page Analyzer reads web pages. Chat mode accepts file attachments.
Yes. HashCoder operates over the full project file tree with file edit and shell tools.
Designs and runs multi-agent pipelines. Define agents, assign roles and tools, connect them in a flow. Voting mode runs all agents independently and compares; chain mode passes output through agents sequentially. When a model fails or rate-limits mid-run, automatic failover routes to a different provider in role-priority order.
Provides a simulated project desktop with a sandboxed file system where an AI agent creates, edits, and organizes files — useful for scaffolding or project planning.
A swarm of models analyzes pasted code or text for malware patterns, trojans, prompt injections, and suspicious logic. Each model votes independently; results combine for a final verdict.
An architecture-first 3D planning agent that produces structured node/mesh plans for game levels, generative architecture, and spatial design.
A full-screen financial analysis studio. Attach a bank statement, CSV, PDF, or XLSX file and receive a structured report with KPIs, transaction table, charts, and recommendations — all computed from your data, never invented.
Ten cloud providers: Anthropic, OpenAI, Google (Gemini), Groq, Cerebras, SambaNova, DeepSeek, Moonshot, Mistral, OpenRouter. Plus Ollama for local models.
Yes, via the Anthropic provider.
Yes, via the OpenAI provider.
Yes, via the Google provider.
Yes, via Ollama. Any Ollama-hosted model works.
Not currently. Local model support is via Ollama only.
Yes. Configure as many API keys as you want and switch freely between models. Agent Swarm mode automatically fails over between providers when one rate-limits or errors.
Only for cloud providers. For local models via Ollama, no API key is required.
For most coding and chat work, frontier cloud models (Claude, GPT, Gemini) significantly outperform local models. For privacy-sensitive or air-gapped use, local models via Ollama are the right tradeoff.
HashCortx is free, open source, local-first, multi-provider, and includes 11 modes beyond coding. Cursor is a paid VS Code fork with a hosted backend focused specifically on AI coding. See Comparison.
HashCortx is a GUI desktop app supporting 10 providers; Claude Code is Anthropic's official CLI for Claude only. HashCortx is broader; Claude Code is deeper for Claude-specific terminal workflows. See Comparison.
HashCortx is a standalone app; Continue is a VS Code/JetBrains extension. Continue lives inside your editor; HashCortx sits beside it. See Comparison.
HashCortx has a GUI and 11 modes; Aider is a terminal-first Git-native pair programmer. Different audiences. See Comparison.
HashCortx is a standalone multi-modal app; Cline is an autonomous coding agent extension for VS Code. See Comparison.
ChatGPT is a hosted web chat app using OpenAI models only. HashCortx runs locally, supports 10 providers, stores no data on a server, and includes 10 specialized modes beyond chat.
Right-click HashCortx → Open → Open. Or run xattr -dr com.apple.quarantine /Applications/HashCortx.app. The v2.0.0 build is unsigned and Gatekeeper blocks it on first launch.
Verify the key in Settings → Providers. Check provider credit/quota. The key is stored in the macOS Keychain; if Keychain Access was reset, re-enter it.
HashCoder requires Permission Guard approval for each new path. Approve via the prompt or the audit modal in Settings.
Confirm Ollama is running and reachable at its default endpoint. Run ollama list in a terminal. Then refresh the model picker in HashCortx.
HashCoder reads the full file tree on project pick. For very large repos, narrow the active project to a relevant subdirectory.
Planned. Not yet released.
Planned. Not yet released.
Planned. v2.0.0 is unsigned.
Modes will be added based on user requests. File a feature request to suggest one.
GitHub Issues for bugs and feature requests. GitHub Discussions for questions and ideas. Releases for shipped updates.
Yes. See CONTRIBUTING.md for dev setup, architecture rules, and how to propose changes.
Bug fixes, performance improvements, additional provider integrations, accessibility improvements, and well-scoped new modes — provided they respect the local-first, no-telemetry, no-backend principles.
Telemetry of any kind. Switching frameworks. Bundlers/transpilers. AI-generated PRs without human review and testing.
Open an issue on GitHub.
See SECURITY.md in the repository for the disclosure process.