Organizes and groups
Sorts what lands by what it actually is — receipts, logins, orders, complaints — so your inbox arrives pre-grouped, not as one undifferentiated pile you triage by hand.
Why it's possible
An agent can't run an inbox it can't reach.
A hosted inbox is sealed. It exposes a thin API over someone else's mailbox, and it holds none of your business — so an agent points at a wall. It can read text. It can't read the order, reach the store, or hold the keys to act.
Tinbox is yours, end to end. The mailbox is legible to software. Your data lives on your account. The keys are yours to lend. So an agent doesn't sit beside the inbox — it operates it: reads the live order, drafts from how you answer, runs the campaign, and parks every send for your one yes.
That's why agent-first is only possible because you own it. You can't bolt an agent onto a rented inbox. You can hand one the keys to your own.
More than a draft button
Stop processing email. Start directing it.
The agent works your whole inbox the way an assistant would — sorting, flagging, grouping, researching, drafting — and answers with cards that deep-link straight into the app, ready to act on. You stay in the loop on anything that sends.
Triages and flags
Surfaces the mail that needs you and quiets the rest. The angry customer gets flagged to the top; the routine confirmation steps quietly aside.
Handles the repetitive
The replies you'd type a hundred times — drafted the moment they arrive, waiting for your yes. The busywork stops being yours to grind through.
Researches and briefs
Ask "what's going on with Acme?" and it traces every thread with that company, summarizes where things stand, and hands you the briefing before the call.
Runs your lists
Feed it your release notes and it drafts the campaign, builds the segment, and schedules the send — the whole lifecycle, right up to your approval. That's the Mailchimp third, run by the same agent.
And the part that needs a human still gets one. The agent does the work; you keep the final say on what sends — the reply, the campaign, the merge.
Skills
Every message arrives ready to act on.
The agent doesn't just sort text. It reaches for Skills — the tools wired to your store, your ledger, your CRM — so an order email renders the live order, a ticket shows the account, and a refund button sits right in the thread. Email stops being a list of messages and becomes an operational surface the agent can act on.
Drive it your way
Bring your own agent. Bring your own model.
MCP
Point Claude, Cursor, or any MCP client at /mcp — one-click OAuth from the Claude app, or a capability-token header — and your agent works the mail from wherever you already are.
CLI
npx tinbox sends, searches, and manages drafts. Every command takes --json and a scoped token, so a script, a cron job, or a coding agent can run it.
SDK
A typed SDK (@tinbox/sdk) for your own apps and automations: send mail, read threads, check delivery — typed against the same API the inbox uses.
The agent isn't trapped behind one chat box. Drive it from the assistant built into your inbox, from Claude or Cursor over MCP, or from your own scripts and workflows — one shared, typed tool manifest behind all of them. And because the keys are yours, you pick the model for each job.
Your keys, the best model per task. A fast, cheap model for autocomplete; a frontier one for the agent. Anthropic, OpenAI, OpenRouter, or Cloudflare Workers AI — no markup, no house model you're stuck with.
Every agent holds a scoped key. Each one authenticates with its own revocable token, checked by the same authorize() gate as every human action. A read-only token can read and draft all day. It can't send.
Autopilot · new
It clears the routine before you sit down.
The co-pilot acts on a turn you start. Autopilot goes further. It runs on a schedule, on its own — triaging, drafting, tagging, tracing — and queues anything that sends for your one-click approval. You set the brief and the tools it may use in Settings. It never crosses the line you draw.
Runs on a schedule
A cron-driven loop wakes on the cadence you set, reads the state of your inbox, and does the reversible work — so the routine is handled before you open your laptop, not only when you ask.
Reversible-only, by construction
Autopilot only runs tools marked reversible — search, draft, tag, summarize. Anything that sends or changes the outside world waits for you, every time. The line is drawn in the code itself, so it holds.
Your approval queue
What it proposes to send lands in a human-in-the-loop queue: approve or reject with a click (or a magic-link), with a TTL so nothing hangs forever. A run history shows what it touched and why it held back.
You hold the dial. Toggle autopilot on or off, rewrite its instructions, and pick exactly which tools it may reach — all from Settings → Autopilot, at runtime, no redeploy. Wire it to your store or billing system and it works your real queue.
Honest about where this is
What ships today, and what's next.
Today
- ✓ Skills organize, classify, flag, and render inbound mail in the background — on every email, with no prompt.
- ✓ The agent reads, traces, researches, summarizes, and drafts across inbox, lists, and sending — on a chat turn you start.
- ✓ Autopilot runs on a schedule, on its own — taking reversible actions (triage, draft, tag) and holding anything that sends for you. Edit its brief and tool list in Settings, no redeploy.
- ✓ Assign an agent to an inbox and it wakes on every inbound email — triaging, tagging, and drafting the moment mail lands, sends held for your approval.
- ✓ You keep the final say on what sends: a one-click approval queue for what the autopilot proposes — the reply, the campaign, the merge.
- ✓ Drive it from chat, MCP, the CLI, or the SDK, with your own model keys — all scoped by capability tokens.
On the roadmap
- → Bounded autonomy: grant the agent a scoped send capability and a budget, and let it handle routine replies you supervise — instead of every send waiting on a human.
- → Per-agent identities and budgets, so support@ and billing@ run as their own scoped principal, not one shared agent.
- → One-click Skill activation from a draft, with fire-count and accuracy telemetry.
The foundations are already here — the autopilot loop, cron, queues, the idempotency ledger, the config-gated tool ceiling, the human-in-the-loop approval queue, and one authorization gate for humans and agents alike.
The inbox that runs itself.
An agent works your queue the way you would — and stops at every line you draw. It's all yours to run: the tool manifest, the agent config, the Skills, on your own account, with your own keys. The lifetime license adds the full readable source. Or run it free first.