Vela
Vela is an agent that lives inside your mail and calendar — reading every thread, holding every commitment, surfacing the one message you half-remember. It drafts and it schedules, but it never sends a word without your say-so.
Autonomous work
This is an agent, not another inbox. It reads, recalls, schedules and drafts — then waits for your word.
Every message and event is indexed the moment it lands — embedded locally on your own hardware, never shipped to a third party. Vector recall and full-text search are fused with reciprocal rank fusion, so Vela finds the thread whether you remember the invoice number or only the argument.
Agent-grade infrastructure for the work in your inbox. Local by default. Approved by you.
Your inbox, understood.
Semantic Recall. 01
- Hybrid vector and full-text search, fused with RRF
- Embeddings run locally — nothing leaves your machine
- Ask in plain language, not Gmail operators
- Re-indexed the second new mail lands
Calendar Command. 02
- Reads freebusy and proposes slots that actually work
- Creates the event and delivers the invites
- Natural-language times resolved to your timezone
Approval Gate. 03
- Nothing sends or invites without your sign-off
- Scripted actions always stop for a human
- Tenant scoped from the session cookie alone
How it works
Four steps from a cold inbox to an agent that remembers. The longest one is the sign-in.
Connect Gmail and Calendar.
One Google sign-in and you're done. Vela never sees a password, and you can revoke its access from your Google account at any moment.
It reads your history, on your hardware.
Every thread and event is embedded locally and indexed. Progress is tracked as it goes, and re-runs skip anything that hasn't changed.
Ask in plain language.
Vector recall and full-text search are fused with reciprocal rank fusion — so you can find the thread by the invoice number, or by the argument you had in it.
Approve before anything leaves.
Drafts, invites and scripted actions all stop at a signed approval gate. Vela does the work; you press send.
Built for
The work you'd never delegate to a filter.
The invoice from the vendor whose name you've forgotten.
Describe what you remember instead of guessing at search operators. Semantic recall runs over everything Vela has indexed, so the thread surfaces from the shape of it — the amount, the argument, the month.
Six people, three timezones, one hour that works.
Vela reads freebusy across the calendars it can see, proposes slots that are genuinely open, then creates the event and delivers the invites — once you've approved the one you want.
The thing you promised three weeks ago and haven't thought about since.
Every commitment is indexed the moment it lands. Ask what you owe people this week and Vela reads back the sentence you wrote, in the thread you wrote it in.
Pricing
Priced per person, not per query. Your embeddings run locally, so there's no usage meter to watch.
Free
See if it remembers what you don't.
forever
- 1 connected inbox
- 30 days of recall
- Embeddings run locally
- Ask anything, in plain language
Pro
The whole agent, one inbox.
per month
- Everything in Free
- Unlimited semantic recall
- Full mailbox backfill
- Calendar Command — freebusy, slots, invites
- Realtime re-indexing as mail lands
- Approval gate on every send
Team
One agent, one shared memory.
per user, per month
- Everything in Pro
- Shared threads and workspace
- Per-seat tenant isolation
- Admin roles and an audit of approvals
- Priority support
Every plan runs embeddings locally. Nothing is sent without your approval.
Questions
The things worth asking before you hand over an inbox.
Your index doesn't. Embeddings are computed on your own hardware with a local MiniLM model, so your mail is never handed to an embedding provider, and the vectors stay in your database. Reasoning is different: Vela's agent runs on Claude, so the message content it needs to answer a given question is sent to the model for that request. Search, ranking and storage are local; the thinking is not.