We built inbox.dog because our B2B SaaS support team was drowning in emails.

We were spending 2-4 hours/day on repetitive tickets (password resets, billing questions, feature requests) while complex issues waited.

Response times could be up to 4+ hours if it needed to pull in a dev to answer, and customers were frustrated.

So we built AI agents that connect to Gmail and automatically sort, reply, and escalate support emails. They understand context, match your team's voice, and only escalate what actually needs human attention.

We cut our response time from hours to minutes, handling 200+ emails/day with 2 people.

How we built it:

   - Frontend: SvelteKit (Svelte 5) with a clean UI for managing agents
   - Backend: Cloudflare Workers + Durable Objects for stateful email processing
   - Database: Cloudflare D1 (SQLite) with Drizzle ORM
   - AI: Two-stage processing (cheap glance filter → full processing) to optimize costs
   - Architecture: Multi-agent system with fan-out pattern (one Gmail fetch → many agents)
   - Auth: Better-Auth with Google OAuth for Gmail API access
   - Tools: 30+ unified tools (reply, forward, label, archive, webhooks, etc.)

Try it (2 minutes): Connect Gmail → pick the Support template → point at a label → Run once with approval → watch Live Log.

   - Custom system prompts - Define the agent's personality and behavior
   - Trigger filters - Gmail query syntax to match emails (e.g., "from:support@ or subject:password")
   - Action instructions - What the agent should do (reply, escalate, label, etc.)
   - Knowledge base - Upload files (docs, FAQs, past emails) that agents search via RAG (vector search) to answer questions accurately
   - Tool permissions - Control which actions each agent can take
   - Scheduling - Cron-like schedules for time-based actions
   - Webhooks - Connect to Zapier, Make, or custom endpoints
   - MCP servers - Connect to other Model Context Protocol servers (HubSpot, Stripe, etc.) so agents can access external data

The agents can read all incoming email, or run on a set schedule.. or you can "run them" on-demand.

Agents use LLMs to understand context and execute actions via the Gmail API.

Privacy/Safety:

   - CASA2 certified. OAuth scopes (gmail.readonly, gmail.modify). 
   - Tokens encrypted at rest. 
   - Live logs and audit trail.
   - Emails are never stored on our servers or databased

Pricing:

   - $99/month per organization (10,000 emails included, then $0.02/email). 
   - Unlimited Agents and Gmail accounts. 
   - 7-day free trial, no credit card. 
   - Processing pauses at your configured limit; agents stay configured.

Try it:

   https://inbox.dog 
   (requires Gmail OAuth, no CC required)

Would love feedback on:

   - The agent configuration UI
   - Whether the two-stage AI processing makes sense
   - What other email actions would be useful
   - How to make it easier for non-technical teams

Happy to answer questions..

Thanks!

   - Jordan Coeyman (x.com/acoyfellow)

Probing Knowledge Holes in Unlearned LLMs