Lauther

io.github.csanadymiklosPublished by GitHub account csanadymiklosv0.1.1Updated Nov 3, 2026

Puts Approve/Deny on your phone before an agent acts, and waits for your fingerprint.

Overview

AI-generated overview

Lauther sends approval, question, location and photo requests from a coding agent to your phone and waits for your answer before the agent proceeds.

What you can do
  • Ask for approval: blocks until you tap Approve or Deny on your phone, optionally behind a fingerprint, and reports APPROVED, DENIED or NO ANSWER.
  • Ask short questions: waits for you to type an answer such as a verification code or a choice, up to 500 characters.
  • Request location: asks you to share one position, returning coordinates, accuracy and read time, or a decline, timeout or expiry.
  • Request a photo: asks for one picture taken on the spot or picked from your library and returns the image plus a link valid for 7 days.
  • Send one-way notifications: pushes a title, message, priority and link, with only urgent ones crossing quiet hours.
When to use it
Use it when an agent runs commands you cannot undo, such as deploying, deleting data, force-pushing or touching production, and you want a human confirmation on your phone first. Also useful for handing short facts, a location or a photo to the agent on request.
Requirements
Runs locally over stdio via npx (Node 18+), nothing to install. Needs a Lauther token in the LAUTHER_TOKEN environment variable, created in the Lauther app under Apps, plus the Lauther mobile app on a registered device and network access to the hosted relay.
Before you install
The service is hosted only and not end-to-end encrypted, so do not put secrets in question or detail text. The token is a credential for your account and should be kept private. Approvals, shared locations, photos and typed answers are sent through a third-party relay, and photo links stay valid for 7 days. Approving lets the agent carry out the action it asked about.

Installation

In SourceWeft

  1. Open Lauther in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

lauther-mcp

Let your coding agent ask you first, on your phone, before it does something you can't undo.

Agent:  About to drop and recreate `users` on staging. Approving?Phone:  [ Approve ]  [ Deny ]        ← optionally behind your fingerprintAgent:  APPROVED after 34s. Confirmed by biometric.

Works with Claude Code, Cursor, Windsurf, or anything else that speaks MCP.

Setup

Get a token in the Lauther app: Apps → + → New token. Then add the server to your MCP config:

json
{  "mcpServers": {    "lauther": {      "command": "npx",      "args": ["-y", "lauther-mcp"],      "env": { "LAUTHER_TOKEN": "lpt_your_token" }    }  }}

Nothing to install — npx fetches it. Node 18+.

The tools

ask_approval

Blocks until you answer, then tells the agent one of three things:

ResultWhat the agent is told
APPROVEDproceed with exactly what was asked, and nothing further
DENIEDstop; do not retry, do not find another way
NO ANSWERstop — silence is not consent

That last row is the one that matters. A timeout is not a yes, and the tool says so in words the model has to read.

Options: details for context, require_biometric to demand a fingerprint, timeout_seconds for how long you're willing to wait (default 120).

On waiting. Lauther's API long-polls for at most 25 seconds. Past that this server keeps asking about the same notification rather than sending another, so a ten-minute wait still puts one question on your phone — not one every half minute.

send_notification

One-way: "the migration finished", "the build broke". Takes title, message, priority and url. Only urgent crosses your quiet hours.

It reports what actually happened, including the case worth knowing about: if no device is registered on your account, you're told that rather than given a cheerful success for a notification nobody received.

get_location

Asks you to share where you are, once. You get a prompt on your phone with Share location and Decline; the agent receives one position, as it was when you tapped.

ResultWhat the agent is told
SHAREDlatitude, longitude, accuracy in metres, and when it was read
DECLINEDdo not retry and do not find the location another way
NO ANSWERsilence is not consent; nothing was shared
EXPIREDthe request lapsed on your phone; nothing was shared

Options: question (what you see on the prompt), details, require_biometric, timeout_seconds (default 120).

request_photo

Asks you for one picture, taken on the spot or picked from your library. The agent gets the link and the image itself, so it can look at it. The link works for 7 days, after which the photo is deleted. If the image cannot be fetched the agent still gets the text and the link.

Same options and the same four outcomes as get_location.

ask_text

Asks you a short question and waits for you to type the answer on your phone: a verification code, a choice, a name, a fact only you know. The prompt shows the question and a field, with Reply and Decline. Answers run to 500 characters, so it is for short things.

ResultWhat the agent is told
ANSWEREDexactly what you typed, in quotes
DECLINEDdo not retry and do not get the answer another way
NO ANSWERsilence is not consent; nothing was sent
EXPIREDthe request lapsed on your phone; nothing was sent

Options: question (default "Answer a question?"), details, placeholder (a hint in the field, such as "6-digit code"), require_biometric, timeout_seconds (default 120).

Telling your agent when to use it

Most models won't reach for an approval tool unless asked. Something like this in your project rules works well:

Before anything irreversible — deploying, deleting data, spending money,force-pushing, emailing real people, touching production — callask_approval and wait. Proceed only on APPROVED.

Suggesting it as a habit

The approval is signed and, when you ask for it, biometric — the answer records that a person confirmed it, not merely that something clicked. That is the difference between an audit trail and a log line.

Related

Worth knowing

  • Hosted only. The relay holds the APNs connection and the identity layer.
  • Not end-to-end encrypted. Don't put secrets in the question text.
  • The app is free on the App Store and Google Play.

MIT

Source: packages/lauther-mcp/README.md at commit 5fce223

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.1.1LatestSep 16, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts