HashCortx AI Workbench
Use models from multiple providers in a local-first desktop workspace for coding, documents, and data.
- Source repo
- Hash-7777/HashCortX
- Stars
- ★ 169
- Last updated
- 3d ago
- License
- MIT
- Primary language
- JavaScript
- FA score
- 73/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A developer who wants an agent to edit a real codebase, review diffs, and run tests.
- Not a fit if
- Teams that require encrypted API key storage
- Organizations that require signed applications
- Source review
- 73/100 · Some gaps
What does this agent do, and when should you use it?
HashCortx is a Tauri v2 desktop workspace built with Rust and vanilla JavaScript, with Chats, HashCoder, Agent Swarm, 3D Forge, ERP, Finance, and Sandbox workspaces. Chats supports twenty-one cloud providers, Ollama, and local model apps that use the common chat format; users bring their own keys or choose a local model. HashCoder works against real project files, presents edits as diffs to keep or undo, and can run tests in its built-in terminal. The app also indexes and searches PDFs and notes on the machine, and connects to business systems such as Odoo, GitHub, Stripe, and Supabase over MCP. It has no HashCortx backend, telemetry, or account; installers are provided for macOS Apple Silicon and Windows, while Linux builds and passes CI but the README says it has not yet been run by anyone.
In Chats, users can choose a model, send messages, compare two model responses, attach files, and query a local knowledge base; the knowledge base splits documents into passages of about 1,200 characters and indexes them by keyword and bge-small-en-v1.5 vectors. HashCoder reads project files, agent instructions, and recent commits, plans the request, proposes file diffs, and can run commands, tests, and browser-like website checks after approval; larger changes may also go through a reviewer step. Agent Swarm builds teams from a canvas, template, or task description, shows a live trace, and exposes each agent's files and versions. 3D Forge turns descriptions into dimensioned models and exports STL, OBJ, 3MF, or STEP; the ERP agent creates a business app exportable as one HTML file; Finance reads statements, CSV, PDF, and XLSX and calculates totals and charts. Sandbox inspects untrusted code for malware, prompt injection, and suspicious logic.
- A developer who wants an agent to edit a real codebase, review diffs, and run tests.
- A user who wants to compare cloud models or switch to local Ollama models in one desktop app.
- A researcher who needs to search PDFs and notes locally without uploading them for indexing.
- A team that wants to assemble task-specific agent groups and inspect their traces and output files.
- A maker who needs dimensioned parts for printing or CAD, or a quick business-record app from a description.
- An analyst who wants to read financial spreadsheets locally or perform an initial inspection of untrusted code.
How do you install or deploy this agent?
Download an installer for macOS Apple Silicon or Windows from the latest release; Linux and Intel Mac users build from source. Source builds require Node 18+, Rust installed via rustup, and platform-specific build tools and libraries. After installation, add and test a provider key under Settings → API keys, or configure a local model under Settings → Local model. The Windows installer omits the embedded search model and uses keyword-only knowledge search. The macOS build is not notarized, so the first launch requires right-clicking the app, choosing Open, and confirming again.
cd ~
git clone https://github.com/Hash-7777/HashCortX.git
cd HashCortX
npm install
npm run tauri buildHow do you use this agent?
Open the desktop app and add and test a key for the chosen provider under Settings → API keys, or configure Ollama or another local model under Settings → Local model. Send a message in Chats; to index files locally, enable the knowledge base in Agents and attach a PDF or text file. In HashCoder, select a project and describe the change; review each diff and choose Keep or Undo, and set the permission mode for command approvals. You can also describe a team task in Agent Swarm, or use the 3D Forge, ERP, and Finance workspaces. Configure an MCP connection to use a supported business system; each change asks for confirmation.
What are this agent's strengths and limitations?
- One desktop app combines chat, a coding agent, agent teams, 3D CAD, ERP, finance, and code security inspection.
- Works with twenty-one cloud providers, Ollama, and local model apps using a common chat format, so users can choose their model path.
- HashCoder presents file edits as diffs with Keep and Undo, built-in terminal tests, and permission modes for command approval.
- The knowledge base indexes by both keyword and meaning on the machine; its search model is built in and works offline from first launch.
- MCP business-system connections keep their keys away from pages and models, and require confirmation for each change.
- API keys are not encrypted; they are protected by the operating system user account.
- The macOS build is not signed or notarized, so first launch requires a manual confirmation step.
- Linux passes CI but the README says nobody has run it yet; the Windows build provides keyword-only knowledge search.
- Cloud models and web search need internet access; requests to a cloud provider include the content sent to the selected model.
- Building from source requires Node, Rust, and platform toolchains; the default embedded-model build also needs a newer AVX2/BMI2 processor.
How does this agent compare with similar options?
Users can choose between cloud providers and local model apps such as Ollama; cloud requests go to the selected provider, while local model tasks can stay on the device. The README does not make direct feature comparisons with other desktop workspaces.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| HashCortx AI Workbench This agent | 73 · Some gaps | Desktop appFree + model costs | ★ 169 | 3d ago | JavaScript | OpenAI API · Claude API |
| agentty | 44 · Major gaps | CLIFree + model costs | ★ 623 | today | C++ | OpenAI API · Claude API |
| OpenClaude — Universal Terminal AI Coding Agent | 42 · Major gaps | CLIFree + model costs | ★ 34k | today | TypeScript | Claude Code · OpenAI API · Claude API |
| Locally Uncensored | 64 · Some gaps | Desktop appFreemium | ★ 2.5k | 1d ago | TypeScript | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The security document details permission modes, Rust path boundaries, a denylist, prompts for network access, and audit logging, supporting relatively strong least-privilege, confirmation, and data-flow scores. Sensitive-data protection is partial because keys are stored in plaintext and most provider requests originate in the renderer. The materials mention dependencies such as ONNX Runtime and CI, but provide no dependency vulnerability audit. Exports, connected systems, and external pages can have external effects, though user-triggered or confirmed. Diffs, Undo, and some atomic writes provide recovery, but the evidence does not justify full marks. Openverse attribution rules and the MIT license are documented; unverified publisher identity does not affect scoring.
The README, package.json, and CI are broadly consistent about the version, checks, and platform targets. CI covers source checks and Rust builds and tests, but its own description says this does not prove the app works; the README says Linux has not been run by anyone yet. Platform prerequisites and the Windows build's omission of the embedding model limit availability. The documents describe some failure cases and messages, but do not show comprehensive user-facing failure handling.
The materials cover chat, coding, knowledge retrieval, swarms, 3D, ERP, finance, and local or cloud models, and spell out tool boundaries, permission modes, and model capability tiers. Many triggers have specific rules, such as confirmation for web searches and local-model modes, but some automatic behavior still performs preliminary searches or provider fallback, and the scope is complex. Platform support is described alongside Linux's lack of hands-on use and older-processor constraints.
The README is organized by features, security, installation, FAQ, and ecosystem, with macOS, Windows, and source installation steps plus platform troubleshooting. The FAQ and known limitations are concrete. The MIT license file is complete. README and package.json agree on the version and link to a changelog, but its contents are not included, limiting assessment of release history and update cadence. The author is named and an issue path is given, but publisher verification and evidence of long-term maintenance responsibility are limited.
The desktop app combines multi-provider chat, a coding agent, agent teams, and several workspaces, with specific features supporting its claimed value. Diffs, Undo, knowledge citations, and swarm traces help users inspect and use outputs. Value and cost depend on the user's hardware, provider charges, and actual behavior of complex features; static materials cannot establish realized benefits or runtime quality, so scores remain conservative.
The security document ties many claims to named source files and commands, and CI describes its check coverage while explicitly saying checks do not prove the app works. README, package.json, the license, and CI corroborate version, licensing, and engineering process. However, the cited implementation files are not included, and many product capabilities are described primarily by the project's own documentation. The materials distinguish described mechanisms and limitations from runtime behavior they cannot establish.
- API keys are stored as a plaintext JSON bundle in app storage rather than encrypted; other processes running as the same user can read them.
- Linux is listed as supported, but the README says nobody has run it yet; passing CI is not runtime validation.
- Cloud models, web search, and connected business systems receive relevant requests or data; review the specific payload and destination before use.