AI Email Assistant

A multi-agent email triage and reply-drafting assistant built with LangGraph + CrewAI on EdgeOne Makers — classifies your inbox, drafts replies with a three-role crew, and lets you approve before anything is sent.

Framework: LangGraph · Category: Orchestration · Language: Python

Overview

AI Email Assistant processes an inbox end-to-end: it fetches emails, classifies them with an LLM, prioritizes by user-defined rules, and drafts replies using a three-agent CrewAI crew (Analyst → Writer → Polisher). Every draft pauses at a human-in-the-loop checkpoint — you approve, edit, reject, or regenerate before the system takes action. The pipeline streams real-time progress to a React UI via SSE.

• Multi-agent drafting — a sequential CrewAI crew (triage analyst, reply writer, voice polisher) produces context-aware, tone-matched replies
• Human-in-the-loop approval — LangGraph interrupt() pauses the pipeline at each draft; resume with approve / edit / reject / regenerate / skip
• Real-time pipeline visualization — SSE streams node-level progress; the UI renders a live flow diagram + streaming narration
• Pluggable email source — ships with 10 realistic mock emails; switch to a live IMAP mailbox with one env var

Environment Variables

Variable	Required	Description
AI_GATEWAY_API_KEY	Yes	Model gateway API key. Use your Makers Models API Key, or any OpenAI-compatible provider key.
AI_GATEWAY_BASE_URL	Yes	Gateway base URL. For Makers Models, use https://ai-gateway.edgeone.link/v1.
AI_GATEWAY_MODEL	No	Model ID. Defaults to @makers/deepseek-v4-flash.
EMAIL_PROVIDER	No	mock (default) or imap. Controls the email data source.
IMAP_HOST	No	IMAP server hostname (e.g. imap.gmail.com). Required when EMAIL_PROVIDER=imap.
IMAP_USER	No	IMAP login username / email address.
IMAP_APP_PASSWORD	No	App-specific password (Gmail: create one here).

This template follows the OpenAI-compatible standard — you can point these variables at Makers Models or any other compatible gateway / provider.

How to get AI_GATEWAY_API_KEY

1. Open the Makers Console.
2. Sign in and enable Makers.
3. Go to Makers → Models → API Key and create a key.
4. Copy it into AI_GATEWAY_API_KEY (set AI_GATEWAY_BASE_URL to https://ai-gateway.edgeone.link/v1).

Built-in models (@makers/deepseek-v4-flash, @makers/hy3-preview, @makers/minimax-m2.7) are free and rate-limited — great for prototyping. For production, bind your own provider key (BYOK) in the console.

Local Development

Prerequisites: Node.js, npm, Python 3.11+

npm install
cp .env.example .env
# Fill in AI_GATEWAY_API_KEY and AI_GATEWAY_BASE_URL in .env
edgeone makers dev

Open http://localhost:8080/agent-metrics for the local observability panel.

Project Structure

email-assistant/
├── agents/email/                   # Backend: Python agent handlers
│   ├── run.py                      # /email/run — main SSE entry (fetch→classify→draft→review loop)
│   ├── review.py                   # /email/review — HITL resume (Command(resume=decision))
│   ├── history.py                  # /email/history — conversation list / get / delete
│   ├── stop.py                     # /email/stop — abort active run
│   ├── health.py                   # /email/health — liveness probe + provider info
│   ├── _graph.py                   # LangGraph StateGraph definition & compilation
│   ├── _state.py                   # EmailAssistantState TypedDict
│   ├── _nodes.py                   # 7 node functions (fetch, classify, prioritize, draft, review, apply, summarize)
│   ├── _routing.py                 # 3 conditional edge functions
│   ├── _crew.py                    # CrewAI crew adapter (builds kickoff inputs)
│   ├── _models.py                  # Pydantic v2 domain models (Email, DraftItem, ReviewDecision, etc.)
│   ├── _providers.py               # EmailProvider protocol + MockProvider + IMAPProvider
│   ├── _events.py                  # CrewAI→LangGraph event bridge (cross-thread)
│   ├── _llm.py                     # LLM client initialization (AI Gateway)
│   ├── _tools.py                   # CrewAI BaseTool implementations (Tone, Template, ThreadContext)
│   ├── _crews/                     # @CrewBase crew definition (YAML agents + tasks)
│   ├── fixtures/                   # 10 mock .eml files + user_rules.json
│   ├── skills/                     # Skill definitions (email-tone, email-templates, triage-rules)
│   └── prompts/                    # System prompts for LangGraph nodes
├── src/                            # Frontend: React + Vite
│   ├── App.tsx                     # SSE state machine + pipeline reducer
│   ├── components/
│   │   ├── ChatLayout.tsx          # Three-column responsive layout
│   │   ├── EmailInboxTree.tsx      # Left: categorized inbox with search/filter
│   │   ├── ConversationStream.tsx  # Center: message timeline + streaming bubbles
│   │   ├── DraftReviewCard.tsx     # HITL approval card (approve/edit/reject/regenerate/skip)
│   │   ├── NodeFlowVisualizer.tsx  # Right: pipeline node status
│   │   ├── EmailDetailDrawer.tsx   # Slide-out email detail panel
│   │   └── HistorySidebar.tsx      # Session history drawer
│   ├── i18n.tsx                    # Internationalization (zh/en)
│   └── historyStorage.ts           # localStorage conversation index
├── edgeone.json                    # Agent runtime configuration
├── requirements.txt                # Python dependencies
└── package.json                    # Frontend build

Files prefixed with _ are private modules — not exposed as public routes by EdgeOne.

How It Works

The agent runs as a session-mode Python runtime under agents/email/. Requests sharing the same conversation_id are routed to the same LangGraph checkpoint, enabling multi-request HITL loops.

Pipeline Flow

fetch → classify → prioritize → [draft → review → apply]* → summarize
                                  ↑_____ regenerate _____|

1. Fetch (/email/run) — pulls emails from the configured provider (mock fixtures or live IMAP), auto-archives senders matching user rules.
2. Classify — a single LLM batch call tags each email with category, priority (0–100), and needs_reply flag.
3. Prioritize — applies VIP-domain boosts and user rules, filters to actionable emails, sorts by priority descending.
4. Draft (per email) — a three-agent CrewAI crew runs sequentially:
   • Triage Analyst: reads the email, produces a structured brief (intent, key points, suggested template/tone)
   • Reply Writer: drafts the reply body using the brief + optional template
   • Voice Polisher: adjusts tone, appends signature, outputs a typed DraftItem JSON
5. Review — calls interrupt(), pausing the graph. The SSE stream emits human_review_required and closes. The frontend renders a DraftReviewCard.
6. Resume (/email/review) — the user's decision (approve / edit / reject / regenerate / skip) is sent back. LangGraph resumes from the checkpoint via Command(resume=decision).
7. Apply — executes the decision (save draft, archive, mark read). Bumps cursor; if more emails remain, loops back to step 4.
8. Summarize — generates a markdown digest of all actions taken.

Key Technical Details

• SSE streaming: stream_mode=["updates", "custom"] — updates drives the pipeline visualizer; custom delivers real-time narration (progress events).
• Conversation ID: passed via the Makers-Conversation-Id request header. The platform's built-in checkpointer (context.store.langgraph_checkpointer) persists graph state per thread.
• CrewAI integration: crew.kickoff() is synchronous — wrapped in asyncio.to_thread(). Events bridged to the async loop via loop.call_soon_threadsafe.
• Timeout: agents.timeout = 1800 (30 min) in edgeone.json to accommodate long multi-email sessions.

Resources

• Makers Agents Documentation
• Quick Start: Agent Development
• Makers Models

License

MIT