Hermes PRD Workbench

Turn raw requirements into reviewable PRDs, technical specifications, test cases, estimates, and interactive prototypes with Hermes Agent.

中文说明 · Quick start · Documentation · Roadmap · Discussions · Security

Watch the 60-second walkthrough · Download the example deliverables · Explore the complete case

Hermes PRD Workbench is an Apache-2.0 self-hosted workspace for product and engineering teams. Upload source material once, preserve the original Agent result, refine a separate working draft, and continue from the same context into reviews and downstream deliverables.

Version 0.4.0-beta.2 is a fresh-install preview for a single instance on a local or trusted network. It has been validated on macOS arm64 and Windows 11 WSL2 with Docker Desktop Linux containers. Linux amd64 and native Windows deployments have not yet been declared supported.

A concrete example: resolve a purchasing-rule conflict

One meeting note says manager approval is enough up to CNY 5,000; a follow-up says anything below CNY 10,000 needs no finance review. A CNY 8,600 workstation takes a different approval route depending on which note a developer follows.

In this fully synthetic case, a real Hermes review identified eight blocking questions. A maintainer then recorded explicit fictional business decisions, generated a four-module PRD, and saved a separate working draft before producing a clickable prototype and engineering handoff drafts. The package includes the review and decisions, PRD files, HTML Demo, test-case XLSX, technical specification, and an assumption-based effort-estimate XLSX.

Read the input materials, results, remaining questions, and verification record. The recording is an edited replay of real browser interactions; 60 seconds is the video length, not model generation time. All materials are synthetic, model outputs need human review, and the static prototype is not a production procurement system. The walkthrough and generated business documents are in Chinese.

What it produces

Mode	Deliverable
Review	Findings, questions, recommendations, and recorded decisions
Generate	Multi-file PRD package plus an editable current draft
Demo	Self-contained single-file HTML demonstration
POC	Static HTML/CSS/JS prototype package and ZIP
Estimate	Explicit-assumption effort estimate and XLSX export
Test case	Multi-file test plan and XLSX export
Technical specification	Architecture, data model, API, and Markdown package

The original generated result stays immutable. User edits and review decisions are stored separately. Failed work is rerun manually, with no silent result repair or automatic retry.

Quick start

Requirements: Docker Engine or Docker Desktop with Compose v2, plus Python 3.10+. On Windows, run every command inside the WSL2 Linux filesystem and enable Docker Desktop WSL2 integration.

git clone https://github.com/YuxForge/hermes-prd-workbench.git
cd hermes-prd-workbench
./workbench setup
./workbench config
./workbench start
./workbench doctor

setup asks for a tool-capable Chat Completions endpoint, model name, model API key, and initial administrator credentials. Secret input is hidden and stored under .runtime/secrets; non-secret configuration is stored in .env. Neither location is committed.

Open http://127.0.0.1:8180 after doctor succeeds. Stop the stack with ./workbench stop; this preserves the database, uploads, and Hermes state volumes.

See the self-hosting guide for upgrades, backup/restore, CORS troubleshooting, and deployment boundaries.

Product views

All screenshots below use synthetic acceptance data.

Generated deliverable	Mobile workflow

Static POC generation	Sandboxed HTML preview

Generated HTML is displayed in an iframe with sandbox="allow-scripts" and no allow-same-origin, including new-window wrappers. Generated JavaScript is never executed in Node VM or Happy DOM. Raw downloads remain separate artifacts.

Architecture and boundaries

Browser ──HTTP──> Nginx/Web ──> Spring MVC control plane ──> MySQL
                                  │
                                  └──> dedicated Hermes Agent runtime
                                           │
                                           └──> operator-selected model endpoint

Hermes makes business conclusions, priorities, and document-organization decisions. The Java control plane validates run/mode, basic JSON and size limits, and minimum artifact structure. Model credentials never enter the browser, and product tools cannot access arbitrary host files, terminals, browsers, networking, memory, delegation, or skill mutation.

The default deployment does not require Nacos, Redis, MinIO, a queue, or a public SaaS control plane. Browser telemetry is disabled by default. Uploaded material is sent to the model service selected by the operator, subject to that provider's data policy.

Documentation

Contributing and support

Questions, setup reports, and ideas belong in GitHub Discussions. Reproducible defects belong in Issues. Please use synthetic materials and never attach credentials, customer documents, complete prompts, or private model responses.

See CONTRIBUTING.md, the Code of Conduct, and the security policy. Private vulnerability reporting is enabled on GitHub.

License

Hermes PRD Workbench is licensed under Apache-2.0. Third-party components retain their own licenses; Hermes Agent is MIT. See NOTICE and the third-party inventory.