# MeshVault

Message your computer. It does the work.

MeshVault is a front door companion: a messaging application with exactly one
contact in it, the machine in your building. You type the way you would text a
colleague. Five agent lanes split the work behind that one thread and come back
with a single answer. Anything that spends, sends, posts or signs stops and
waits for your yes.

MeshVault is in early access. Founding member installs are open now.

## What is in the box

- The companion. A messaging app rather than a bot in a rented inbox. No phone
  number, no cloud mailbox, and no third party messenger in the middle of the
  thread. The only contact in the app is your own machine.
- The mesh. Devices pair by scanning a QR code and approving it on the machine
  itself. What you end up with is a private network with no public ingress and
  no third party relay sitting in the middle of your work.
- The browser. A real logged-in session under your control, so an agent works
  inside the accounts you already pay for instead of asking you to paste
  screenshots at it.
- The vault. Your existing drives pulled in during the install and written as
  plain Markdown on your own disk, indexed so an agent finds one note instead
  of loading the whole folder.
- The library. 4,418 written procedures an agent loads on demand, one at a
  time, hash checked at fetch.

## Proof you can run yourself

4,418 usable distinct skills, measured 2026-07-27 at 03:29Z. Run this from the
library root:

    find . -mindepth 3 -maxdepth 3 -name SKILL.md -type f -printf '%h\n' \
      | sed 's|.*/||' | sort -u | grep -vc skillspector
    4418

The rule is distinct skill directory names at depth two, with the adversarial
test fixtures excluded. Counting every SKILL.md at any depth returns 13,477,
because that walk counts a skill's own reference files as skills. Counting
directories without deduplicating returns 5,255, because roughly 800 skills are
mirrored across two roots. Both are bigger. Neither is honest.

On cost structure: ICONIQ's survey of AI companies puts inference at 40 to 50
cents of every revenue dollar and reports the trend rising. OpenAI's gross
margin moved from about 40 percent to about 33 percent. Anthropic landed near
40 percent, under its own forecast. The model in a MeshVault install runs on
your hardware, so what your usage costs us is your electricity.

## The hard part

You are handing an agent a logged-in browser.

- The gate is structural. Outward actions cannot proceed without your yes, and
  there is no setting that switches it off.
- Access is scoped and revocable. Credentials stay inside the connector layer
  and never enter an agent's context.
- The node is yours. At handover you get the install record, including command
  output showing what is listening and on which interface.
- Nothing leaves the building while no cloud model is connected. You can
  connect one, some jobs are better with one, and that switch is yours.

If we disappear you keep the hardware and the files. The vault is plain
Markdown on your own disk, so exit is copying a folder.

## How to start

1. Twenty minute walkthrough. You watch the approval gate stop a real outward
   action. No card, no trial clock.
2. A paid audit, then a written scope in your hands within five business days.
3. If the five jobs do not clear the bar, we say so and do not take the
   install. You keep the document.
4. Install day. Mesh paired, vault built, front door live, one named job walked
   end to end with you.
5. All five jobs walked end to end within thirty days, on a recorded call.

Installs are booked one at a time until a second installer is trained.

Available today, separate from the install: the Agent Skills Starter Pack, $49,
one payment. 35 files. Five agent skills, five memory templates, four runbooks,
three routing recipes and four worked examples, as editable Markdown.

The Mac and iPhone apps are in early access. You install on your Mac directly,
with no App Store middleman.

FireDev LLC dba MeshVault · Orange County, California · contact@meshvault.ai
