🏭 Agent Mill — Feature Showcase
This document walks through every major feature of Agent Mill with screenshots and explanations.
1. Command Center — Agent Workstation
The command center provides a real-time overview of all your AI agents, their status, and quick access to conversations and tasks.
| View | Description |
|---|---|
| Grid View | Card-based layout showing agent names, descriptions, status indicators |
| Station View | Visual office-desk diorama with dual monitors, status LEDs, and desk accessories |
Key actions: Start a conversation, create tasks, monitor agent health.
2. Multi-Tab Agent Editor
Each agent is configured through a professional 4-tab editor:
Tab 1: Basic Info
- Avatar upload, unique code, display name, business description
- Enable/disable toggle, set as default agent
Tab 2: Role & Prompt
- Persona setting — Define the agent’s identity, responsibilities, and tone
- Output constraints — Format rules, compliance limits, tone guidelines
Tab 3: Model & Compute
- Model selection (Claude, DeepSeek, Qwen, GLM, OpenAI)
- Compute slider — 5 levels from Lite to Max, with cost/efficiency guidance
- Advanced parameters: Temperature, max tokens, max turns
Tab 4: Knowledge & Skills
- Knowledge base binding — Multi-select RAG knowledge bases
- Skill mounting — Tag-based skill management with quick-add from skill market
- MCP connector binding — Attach external tool connectors
3. Dual-Path Streaming Engine
Agent Mill’s core technical differentiator — two streaming paths that produce identical SSE events:
| Provider | Streaming Mechanism |
|---|---|
| Anthropic (Claude) | Claude Agent SDK native SSE — true streaming with thinking + tool_use events |
| OpenAI-compatible | Custom tool_calls multi-turn loop (DeepSeek, Qwen, GLM, OpenAI) |
SSE Event Order: meta → thinking → text → tool_use → tool_result → file → ui → error → done
What This Enables
- Real-time token streaming — No waiting for full response
- Thinking display — Collapsible “thinking card” showing Claude’s reasoning
- Tool execution trace — Live step-by-step tool call visualization
- Widget rendering — Inline SVG charts, HTML forms, data tables
- File generation — Real-time file creation and preview
4. Skill System (Three Forms)
| Type | Description |
|---|---|
| Path | ZIP upload with SKILL.md + resource files. Ideal for complex business processes. |
| Callable | Python function direct import. Lightweight tools, rapid development. |
| Composite | YAML DAG step orchestration with conditional branching and parallel execution. |
Security
- 9-rule shell injection pattern detection
- Python AST static analysis (blocks eval/exec/subprocess)
- Upload interception with audit logging
5. MCP Connector System
Three transport protocols to connect any external tool:
| Protocol | Use Case |
|---|---|
| stdio | Local subprocess tools (CLI, scripts) |
| sse | Server-Sent Events for long-running connections |
| http | Streamable HTTP for REST APIs |
Management: CRUD + connection test + tool list discovery + auto-generated Chinese summaries + per-agent isolation.
6. RAG Knowledge Base
| Stage | Technology |
|---|---|
| Upload | 30+ file formats (PDF, DOCX, XLSX, images, code) |
| Parse | MinerU engine with auto-fallback |
| Chunk | 500-char window, 50-char overlap |
| Embedding | OpenAI-compatible API (text-embedding-3-small) |
| Vector Store | ZVec (Alibaba, Rust kernel, embedded — no extra containers) |
| Retrieval | Cosine similarity, configurable top_k |
Auto-injection: Every conversation turn automatically retrieves relevant knowledge and injects it into the system prompt.
7. Multi-Agent Collaboration
| Mode | Description |
|---|---|
| Sequential | Agent1 → Agent2 → … → AgentN |
| Parallel | All agents process simultaneously |
| MapReduce | Individual processing → aggregation |
Communication: Asynchronous message passing with priority support (normal/high), message delegation, and auto-consumption polling.
8. Workflow & Solution Pack Editor
- Vue Flow canvas — Drag-and-drop workflow editor
- 6 node types: skill, parallel_group, aggregator, condition, sub_agent, human_approval
- Persistence: Each node writes to DB on execution
- Resumability: Breakpoint recovery for failed workflows
- Human approval: Pause and wait for manual approval
- Execution trace: Complete run history
9. Dashboard & Analytics
- 7+3 ECharts charts: Token trend, user top 10, agent distribution, model cost, department stats, agent effectiveness, latency trend, error rate trend, system health
- Token detail: Paginated query with user/agent/model/date filtering, CSV export
- 4-tab layout: Overview, users, cost, performance
10. Enterprise Features
| Feature | Description |
|---|---|
| RBAC | Three roles: admin/operator/user |
| SSO/LDAP | LDAP bind + OIDC full callback flow |
| Audit Logs | Dual-table: admin operations + call logs |
| Data Masking | Auto-mask email/phone/ID card in logs |
| Rate Limiting | Token bucket per-path |
| Quota Control | User-level monthly token/cost limits |
| Alert Notifications | Rule engine + DingTalk/WeChat/Feishu webhooks |
11. Mobile App
Independent Vue 3 application at /m.html with:
- Hash-based routing (fully separate from desktop)
- Dedicated stores, views, and styles
- Core chat, task management, notification, and space features
12. File Processing
| Format | Engine |
|---|---|
| TXT/MD/CSV/JSON/HTML/XML/YAML | Direct read |
| MinerU → pypdf fallback | |
| DOCX | MinerU → python-docx fallback |
| XLSX | MinerU → openpyxl fallback |
| Images | MinerU OCR |
Preview: HTML iframe, PDF native, Markdown, code, SVG, images (JWT-secured with expiring tokens).
Complete Module Checklist
| # | Module | Status |
|---|---|---|
| 1 | Dual-path Streaming (Anthropic + OpenAI) | ✅ |
| 2 | Skill System (path/callable/composite) | ✅ |
| 3 | MCP Connector (stdio/sse/http) | ✅ |
| 4 | RAG Knowledge Base | ✅ |
| 5 | DAG Solution Pack | ✅ |
| 6 | Workflow Editor (Vue Flow) | ✅ |
| 7 | Template Market | ✅ |
| 8 | Memory System | ✅ |
| 9 | Context Compression | ✅ |
| 10 | Self-Learning | ✅ |
| 11 | Multi-Agent Collaboration | ✅ |
| 12 | RBAC | ✅ |
| 13 | SSO/LDAP | ✅ |
| 14 | Audit & Compliance | ✅ |
| 15 | Data Masking | ✅ |
| 16 | API Rate Limiting | ✅ |
| 17 | Quota Control | ✅ |
| 18 | Alert Notifications | ✅ |
| 19 | Dashboard & Analytics | ✅ |
| 20 | File Processing | ✅ |
| 21 | Security (7 layers) | ✅ |
| 22 | Mobile App | ✅ |