Content
<div align="center">
<img src="desktop/assets/app-icon.png" width="96" alt="ZMT AI Media Agent Logo">
# Tool List
### Packing a Content Team into a Windows Computer
From IP positioning, hit topic selection, copywriting scripts, picture storyboards, to AI videos, material precipitation, matrix publishing, POI loading and review reflux, a production line completes.
[](LICENSE)
[](https://github.com/rfdiosuao/zmt-ai-media-agent)
[](https://www.electronjs.org/)
[](https://go.dev/)
[](docs/ZMT_CLI_MCP_USAGE.md)
**Local Priority · Replaceable Model · Recoverable Task · Published with Evidence · Callable Agent**
</div>
## What Problem Does It Solve
Many AI content tools only complete one button: write copy, draw pictures or generate videos. The part that really takes time is to connect the dozen steps and continue after any step fails.
ZMT AI Media Agent breaks down the content team's work into an executable, savable, and recoverable production line:
```text
Product File -> IP Image -> Account Planning -> Hit Topic Selection -> Hit Copywriting
-> Opening Hook -> Shooting Script -> Storyboard -> AI Video
-> Material Assets -> Release Pre-inspection -> POI -> Release Queue -> Data Feedback
```
It's not a chat box shell, nor does it promise "click once must be a hit". What it really provides is a content production infrastructure that can see progress, precipitate assets, manually take over, and continuously review.
## Core Capabilities
| Module | Capability |
| --- | --- |
| IP Director Pipeline | 10-step guided workflow, automatic saving, breakpoint recovery, failure retry, real-time product preview |
| IP and Customer Assets | Product file, image positioning, target audience, expression boundary, copy style, account strategy |
| Hit Content Production | Topic variation, opening hook, short and long copywriting, structure imitation, shooting script, compliance check |
| Multi-model Adaptation | OpenAI compatible text interface, image interface, small cloud quail video, Seedance and other adaptation layers |
| Image and Video | Picture prompt words, reference materials, image to video, text to video, multi-segment result archiving and de-duplication |
| Material and Script Library | Precipitate topic, copywriting, storyboard, picture, video and release draft by customer, IP, type and status |
| Matrix Publishing | Multi-platform account, release pre-inspection, task queue, failure classification, manual takeover, release evidence |
| POI Workflow | Location search, loading status, pre-release verification, prohibit false report when not clearly selected |
| Agent Access | ZMT CLI, MCP Server, OpenClaw Skill, let Codex, Claude and other Agent call capabilities |
| Local Data | Go Core + SQLite, save business data on the machine, browser Profile isolate account login state |
## Why It's Not Just a Demo
- **State Machine Instead of Loading Circle**: Each step has a clear state, time-consuming, progress and next action.
- **Idempotence and Recovery**: Repeated clicks, network timeouts and application restarts will not easily cause repeated deduction or repeated products.
- **Evidence Priority**: Release success must be confirmed by task status, result link or evidence record, cannot take "clicked the button" as success.
- **Model Decoupling**: Text, image, video providers access through adaptation layer, not bind business process to single supplier.
- **Human-computer Collaboration**: Verification code, platform risk control and final release and other high-risk actions retain manual confirmation entrance.
- **Agent Native**: Desktop UI, CLI and MCP share the same business capability, no need to rewrite integration for each Agent.
## System Architecture
```mermaid
flowchart LR
U["Operator / Agent"] --> UI["Electron + React Workspace"]
U --> CLI["ZMT CLI / MCP Server"]
UI --> IPC["Controlled IPC"]
CLI --> CORE["Go Core"]
IPC --> CORE
CORE --> DB["SQLite Business Data"]
CORE --> JOB["Task, Lease, Idempotence, Recovery"]
CORE --> BROWSER["Isolated Browser Profile"]
JOB --> MODEL["Text / Image / Video Adaptation Layer"]
BROWSER --> PLATFORM["Douyin / Xiaohongshu / Kuaishou and other platforms"]
MODEL --> ASSET["Material and Script Library"]
ASSET --> PUBLISH["Pre-inspection / POI / Release Queue"]
PUBLISH --> FEEDBACK["Evidence / Log / Data Feedback"]
FEEDBACK --> JOB
```
## Quick Start
### Environment Requirements
- Windows 10/11 x64
- Node.js 22 or higher
- Go 1.24
- npm
### Installation and Startup
```powershell
git clone https://github.com/rfdiosuao/zmt-ai-media-agent.git
cd zmt-ai-media-agent
npm install
npm run pw:install
npm run dev
```
After the first startup, complete:
1. Configure text, image and video models.
2. Create customer, product file and IP image.
3. Enter IP director pipeline to generate complete content assets.
4. Confirm pictures, videos and release copywriting in the material library.
5. Add platform accounts and manually complete login.
6. Execute pre-inspection on the release page, confirm POI and target account.
7. Manually confirm and create release tasks, and view results in task logs.
All API keys should be configured through local settings page or environment variables. The repository does not contain any available keys.
## Agent, CLI and MCP
```powershell
npm run zmt -- --help
npm run zmt:mcp
npm run zmt:mcp:config
npm run zmt:mcp:smoke
```
Detailed access methods:
- [ZMT CLI and MCP Usage Guide](docs/ZMT_CLI_MCP_USAGE.md)
- [Agent Access Prompt](docs/ZMT_AGENT_INTEGRATION_PROMPT.md)
- [OpenClaw Agent Guide](docs/OPENCLAW_AGENT_GUIDE.md)
## Build Windows Installer
```powershell
npm run build
npm run dist
```
The installer outputs to `artifacts/nsis/` by default. Before formal distribution, configure Windows code signing certificate and review FFmpeg, model service and third-party platform licenses and service terms.
## Test and Quality Gate
```powershell
go test ./...
npm run test:helper
npm run test:zmt
npm run test:openclaw
npm run build:renderer
npm run test:commercial-smoke
npm run check:commercial-quality
npm run check:commercial-readiness
```
The project includes unit tests, static UI tests, director pipeline visual smoke, commercial quality samples, garbled code scanning, packaging resource verification and sensitive information scanning.
## Data and Security Boundary
- Platform passwords will not be written to SQLite.
- Cookie, Storage and login state are saved by local isolated Browser Profile.
- Verification code and platform risk control require manual processing, no bypass mechanism provided.
- Automatic release must be explicitly confirmed by the user; task creation does not equal platform release success.
- Do not submit `.env`, database, browser Profile, customer materials, API keys or authorization private keys.
- Platform page updates may cause automation adapter to fail, encounter failure should stop and check evidence, should not retry indefinitely.
## Project Structure
```text
desktop/ Electron main process and React interface
cmd/popagent-core/ Go Core entrance
internal/core/ Data, task, director, release and authorization service
tools/agent-helper/ Text, image, video and director Skill
tools/playwright-helper/ Platform browser adapter
scripts/ Build, test, CLI, MCP and delivery scripts
docs/ Architecture, development and Agent access documents
fixtures/ Offline quality verification samples
```
## Roadmap
- [x] 10-step IP director pipeline and automatic saving
- [x] Material/script asset loop
- [x] Small cloud quail and Seedance video adaptation layer
- [x] Release pre-inspection, POI, task queue and failure evidence
- [x] ZMT CLI, MCP and Agent access
- [ ] More model supplier plugins
- [ ] Visual workflow orchestrator
- [ ] Team collaboration and permission system
- [ ] Data feedback and content effect scoring
- [ ] Digital human and automatic clipping plugin
## Participate and Contribute
Welcome to submit Issue and Pull Request. Please read [CONTRIBUTING.md](CONTRIBUTING.md) and [SECURITY.md](SECURITY.md) before starting.
Prioritize contributions: model adapters, legally compliant platform adapters, content quality evaluation, Windows stability, and clearer and less anxious product experience.
## Open Source License
This project adopts [GNU Affero General Public License v3.0](LICENSE). When providing services to users through the network after modification, you also need to provide the modified source code.
Douyin, Xiaohongshu, Kuaishou, WeChat, OpenAI, Seedance, small cloud quail and other names and trademarks belong to their respective rights owners. This project has no official affiliation or endorsement with these platforms.
<div align="center">
**Let AI not only generate content, but truly participate in a deliverable content production line.**
</div>
