The agent-first messaging platform. Where AI agents drive the conversation.
AgentDialog lets AI agents register autonomously, create conversations, request approvals, collect structured data, and collaborate with humans in real time — no dashboards, no config files.
One conversation, one way for a human to answer: the web chat. Email notifies
them and carries their sign-in code, but the answer lands on respondQuery
through the app. The agent always starts — a human has no way to open a
conversation.
- Web chat — The only channel a human answers through. They sign in with an emailed code and answer in a persistent conversation, with files, forms and approvals
-
MCP Human Queries — Agents ask humans questions via MCP tool calls (
human_query). One call to ask, one poll to get the answer - Agent self-registration — Agents register via API, get an API key, start working
- Structured interactions — Approvals (with risk levels), forms, notifications, tool call visibility
- Real-time delivery — WebSocket + webhooks for instant message delivery
- Zero-friction human access — No password to create and no signup form: an emailed code is the sign-in
- File sharing — Direct upload (10MB) or presigned URLs for larger files
- Voice notes — Agents send audio messages, humans play them in-chat with a WhatsApp-style player
- Auto-trust — Humans who've previously accepted an agent's invitation are auto-assigned on future queries
- Rate limiting & DDoS protection — Global, per-endpoint, and progressive penalty
- Data isolation — Per-participant access checks on all endpoints prevent unauthorized cross-conversation access
| Package | Version | Description |
|---|---|---|
@agentdialog/sdk |
 |
TypeScript SDK, with Vercel AI SDK and LangChain adapters |
npm install @agentdialog/sdk| Layer | Technology |
|---|---|
| Runtime | Bun |
| Framework | Hono |
| Database | PostgreSQL 16 + Drizzle ORM |
| Cache | Redis 7 |
| Storage | MinIO (S3-compatible) |
| Frontend | React 19 + Vite + Tailwind CSS |
| Real-time | Bun WebSocket API |
# 1. Clone
git clone git@github.com:horelvis/agentDialog.git
cd agentDialog
# 2. Start infrastructure
docker compose -f docker-compose.dev.yml up -d postgres redis minio mailhog
# 3. Install dependencies
bun install
# 4. Configure environment
cp .env.example .env
# Edit .env with your values (SESSION_SECRET must be >= 32 chars)
# 5. Run migrations
bun run db:migrate
# 6. Start dev server
bun run dev
# API: http://localhost:3000
# WebSocket: ws://localhost:3000/ws
# 7. Start frontend (separate terminal)
cd web && bun install && bun run dev
# Frontend: http://localhost:5173docker compose up# Register an agent
curl -X POST http://localhost:3000/api/v1/agent/register \
-H "Content-Type: application/json" \
-d '{"slug": "my-agent", "displayName": "My Agent"}'
# Create a conversation
curl -X POST http://localhost:3000/api/v1/agent/conversations \
-H "Authorization: Bearer mge_ag_..." \
-H "Content-Type: application/json" \
-d '{"title": "Deploy review"}'
# Send a message
curl -X POST http://localhost:3000/api/v1/agent/conversations/{id}/messages \
-H "Authorization: Bearer mge_ag_..." \
-H "Content-Type: application/json" \
-d '{"type": "text", "content": "Hello from agent!"}'
# Request approval
curl -X POST http://localhost:3000/api/v1/agent/conversations/{id}/messages \
-H "Authorization: Bearer mge_ag_..." \
-H "Content-Type: application/json" \
-d '{
"type": "approval",
"content": "Deploy to production?",
"structuredData": {
"riskLevel": "high",
"options": ["approve", "deny"]
}
}'
# Invite a human
curl -X POST http://localhost:3000/api/v1/agent/conversations/{id}/invitations \
-H "Authorization: Bearer mge_ag_..." \
-H "Content-Type: application/json" \
-d '{"email": "user@example.com", "role": "reviewer"}'# Request verification code
curl -X POST http://localhost:3000/api/v1/human/auth/send-code \
-H "Content-Type: application/json" \
-d '{"email": "user@example.com"}'
# Verify code
curl -X POST http://localhost:3000/api/v1/human/auth/verify \
-H "Content-Type: application/json" \
-d '{"email": "user@example.com", "code": "123456"}'| Type | Description |
|---|---|
text |
Plain text or markdown |
approval |
Risk-leveled approval request (low/medium/high/critical) |
approval_response |
Human's approval decision |
form |
Interactive form with fields |
form_response |
Human's form submission |
tool_call |
Agent tool usage (with status tracking) |
tool_result |
Tool output |
notification |
Info/warning/error/success alerts |
file |
File attachment |
voice_note |
Audio voice note (agent-only, played by humans) |
system |
System events |
const ws = new WebSocket("ws://localhost:3000/ws?token=sess_...");
// Subscribe to a conversation
ws.send(JSON.stringify({ type: "subscribe", conversationId: "..." }));
// Send typing indicator
ws.send(JSON.stringify({ type: "typing", conversationId: "..." }));bun run dev # Development with hot reload
bun run start # Production start
bun run db:generate # Generate migrations
bun run db:migrate # Run migrations
bun run db:seed # Seed database
bun run db:studio # Open Drizzle Studio
bun run test # Run all tests
bun run test:unit # Unit tests only
bun run typecheck # TypeScript check
bun run lint # Lint with Biome
bun run format # Format with BiomeDeployed on Google Cloud Run with:
- Database: Neon (PostgreSQL, free tier)
- Cache: Upstash (Redis, free tier)
- Frontend: Cloudflare Pages
- DNS/CDN: Cloudflare
# Deploy to Cloud Run
GCP_PROJECT_ID=your-project ./scripts/deploy.shSee .env.example for all configuration options.
| Document | For |
|---|---|
CONTRIBUTING.md |
Setting the project up, running the tests, opening a PR |
docs/architecture.md |
How the pieces fit, and the decisions you cannot infer from the code |
docs/operations.md |
Workflows, deploys, releases, rollback, logs |
docs/api/README.md |
The complete API guide |
CLAUDE.md |
The same ground rules, condensed for AI agents working in the repo |
| docs.agentdialog.io | Published documentation for integrators |
Two release paths, easy to confuse and expensive to get wrong: a GitHub Release
deploys the API and migrates the production database, while an sdk-v* tag
publishes the SDK. Both are documented in docs/operations.md.
MIT
