What MeshVault is,
and what is real today.
MeshVault is a workspace daemon you run on your own machine. It holds your chats, your files and your search index on that machine's disk, talks to a model on that same machine, and stops for your approval before it sends, spends or writes.
One process, bound to loopback.
The workspace is a Next.js server started with -H 127.0.0.1. It listens on port 3000 of the machine you started it on and nowhere else. Nothing in the repository opens a network interface. Checked on the pilot node while writing this page:
$ ss -ltnp | grep -E ':3000|:11434'
LISTEN 0 511 127.0.0.1:3000 0.0.0.0:* users:(("next-server (v1",pid=506691,fd=22))
LISTEN 0 4096 127.0.0.1:11434 0.0.0.0:*
Port 11434 is Ollama. That is where prompts, transcripts and embedding requests go, which is the reason there is no inference bill to pass on to you. Both addresses are loopback, so both are reachable only from that machine.
Four runtime dependencies: next, react, react-dom, playwright-core. No database, no vector store, no Stripe SDK. Durable state is JSON files under ~/.meshvault/workspace/store/ plus a binary Float32 index under ~/.meshvault/workspace/index/.
What works, and where to read it.
Each row was read out of the source on 2026-07-27. Where a page here goes deeper, it is linked.
| Capability | State | Where it lives |
|---|---|---|
| Passphrase gate and sessions | Working. scrypt with a 16-byte salt, sessions valid 30 days, tokens stored as SHA-256 only | lib/auth.ts |
| Mesh identity and QR pairing | Working. Ed25519 keypair generated on the node, owner-approved device enrolment | Getting started, Security |
| Approval gate | Working, with one gap named on its own page | Approvals |
| Skill retrieval | Working. 4,418 skills on disk, reached through a router, one body at a time | Skills |
| Local models | Working, through Ollama on 127.0.0.1:11434 | lib/ollama.ts |
| Browser arm | Working. Headless Chromium driven by playwright-core; every mutating action needs an approval | Privacy |
The list we would rather you read first.
Being wrong in the flattering direction is still being wrong. Every item here is a thing the product could be read as claiming and does not do.
The mesh carries no traffic
lib/mesh.ts generates a keypair, mints pairing invites and records devices that an owner approved. It establishes no tunnel, assigns no addresses and routes nothing.
A device in that ledger is enrolled. Treat it as a record of a decision you made, not as a live link.
Plans are not open
Billing mode in the default settings is pilot. The Stripe code refuses outright to create a checkout session with a live key, so the pilot cannot take real money even by accident.
The $49 skills pack on the home page is the one thing you can buy today.
iMessage is off the table
A bridge exists in the tree and has never carried a message on the pilot node. Its two routes are not in the middleware's open-path list, so an external watcher gets a 401 before any token check runs.
Apple's Messages for Business terms require a live human agent in business hours, which rules out the version people picture.
Connectors are a policy, not a wire
The Connectors page states that Composio is the only external-account layer and the settings object carries a composioOnly flag. No code in the application calls Composio.
MCP servers you register are written to a settings file. No client connects to them in this version.
L1 and L2 are the same
The approval levels are documented as "ask once" and "ask every time". The gate has no branch on L1 and nothing records that a tool was approved once, so choosing L1 changes nothing.
Full truth table on the approvals page.
The apps are scaffolds
The iPhone and Mac apps are in design and scaffold phase. Neither is downloadable, neither is on the App Store, and nothing on this site should be read as saying otherwise.
The workspace itself installs from source on a machine you already own.
The economics, stated once.
Hosted assistants pay for inference on every request they serve, and that cost is rising as a share of revenue rather than falling. OpenAI's gross margin moved from 40% to 33%. Anthropic missed its own forecast at 40%. ICONIQ's read of the sector points the same direction.
A model running on your machine moves that cost to your electricity bill, where it is already partly sunk. That is the whole argument. It is also why the pairing model matters: if the work happens on hardware you own, the only question left is who is allowed to ask it to work.
The known failure mode of this design is hardware gating. Microsoft spent twenty months requiring an NPU for Copilot+ features and dropped the requirement at Build 2026. Rosetta 2, Half-Life 2's DirectX 7 path and Apple's Private Cloud Compute all shipped a degraded fallback for the machines that could not run the fast path. Crysis, Vision Pro and Humane did not. We would rather be in the first list.
Four pages, in order.
- Getting started. The real first run: set a passphrase, create a mesh, scan the QR, approve the device.
- Approvals. What L0, L1, L2 and L3 permit, tool by tool.
- Skills. The library, how retrieval picks from it, and how the count is arrived at.
- Privacy and Security. What leaves the device, and why a photograph of the pairing screen is worth nothing to an attacker.
These pages describe software behaviour. The legal privacy policy for meshvault.ai and the skills pack is at /privacy, and the terms are at /terms.