Newmetor memory takes #1 on EnterpriseRAG-Bench

A home for your agents.
On your own subscription.

metor runs delegable bots on a machine you own. Each bot has a name, a memory, its own browser and terminal, and a schedule of its own, driven by the official Claude Code or Codex with your login. Open source, one-line install.

$curl -fsSL https://www.metor.com/install.sh | sudo bash

A fresh Linux server, five minutes, no domain required.

A home, not a chat window

Most agents are sessions. A bot in metor stays.

You open a window, get an answer, close it, and everything is gone. Not here.

It has a name and a role.

You create a bot with a name and a role in one or two sentences: what it is for and how it should work. The role becomes its standing instructions. Tasks come later, in the chat, one at a time or as a schedule.

It remembers.

A bot's files, history and session context live in its own directory. Restart the machine and the bot comes back with the context it had. Pause it for a week and it picks up where it stopped.

It has its own browser and terminal.

Every bot gets its own screen with a browser and a terminal, and its own browser profile: sign in to a site once for that bot, and the login stays. Watch the screen at any time and take over with mouse and keyboard when a page asks for a login, a second factor or a captcha.

It works while you are away.

Put metor on a server and the bots keep working when your laptop is closed and you are asleep. Routines fire on schedule, long tasks run to the end, and what a bot found is waiting in the chat in the morning.

How it works

One machine, many bots.

metor runs as one Linux container on a machine you own: a server, an old PC, your Mac. Inside it, every bot has a directory, a session of its runtime, and its own desktop: a virtual screen with a browser window and a terminal. A gateway serves the interface to your phone and to your desktop browser.

one Linux container · your machine
atlasresearches and writes digests
directorysessionbrowserterminal
ledgerkeeps the books
directorysessionbrowserterminal
scoutwatches three feeds
directorysessionbrowserterminal
gateway · port 6010phonelaptopwatch link

Chat

Answers stream in as Markdown. Grey tool lines show what the bot is doing, expandable to the full command and its result. When the runtime asks for permission, an approval card appears: allow or deny. A stop button interrupts a running turn; the bot keeps its context.

Computer panel

Next to the chat: the bot's screen, a real terminal in the bot's directory, and a file browser with downloads. Drop files into the chat or paste screenshots; the bot returns files as cards with a preview or a download.

Phone and desktop

The same interface follows the messenger pattern on a phone: the bot list, then the chat with a back gesture. Side by side is for desktops. No app to install.

Bring your own subscription

You already pay for Claude or ChatGPT. metor does not add a bill.

Every bot is a session of the official coding-agent harness, Claude Code or Codex, signed in with your own subscription through the official login flow. No API keys, no tokens copied out of a config file, no proxy in between. The same harness you use at your desk, now running on its own.

Claude Coderuntime

your Claude subscription

OpusSonnetHaiku
Codexruntime

your ChatGPT subscription

GPT models
OpenCodeon the roadmap

GitHub Copilot, ChatGPT

roadmap
Claude quota · 5h16% · week 18%

Shared by all Claude bots. When it is used up, Claude bots wait for the window to reset; Codex bots keep going.

Runtime and model per bot.

Claude Code with Opus, Sonnet or Haiku; Codex with the current GPT models. Mix them freely across bots.

Sign in from the interface.

The “New bot” dialog runs the official login of the runtime: a link, a code, done. Once per machine, not per bot.

Quota in view.

The sidebar shows the Claude quota for the five-hour and the weekly window; all Claude bots share it.

Runtimes are a seam, not a lock-in.

A registry defines what a runtime is. Claude Code and Codex are the two today; OpenCode is next on the roadmap.

Routines

There is no form for schedules. You tell the bot.

“Every weekday at 7:00, check the front page again and write a fresh digest.” The bot creates the schedule itself and confirms it. A panel lists every routine with its next run; changing, pausing or deleting one is a sentence in the chat.

Missed runs are caught up.

A run that was missed while the machine was down runs once after the restart, not twenty times.

Auto-pause.

A routine that has run twenty times without you saying anything pauses itself and tells you so. A schedule nobody looks at should not spend your quota.

Runs arrive in the chat.

Each run is a grey “Routine” message followed by the bot's work, so the history stays one conversation.

Your machine

A server, an old PC, your Mac. One container.

root@server
$ curl -fsSL https://www.metor.com/install.sh | sudo bash
metor – server installation
Where should metor be reachable?
Address [188-68-58-3.sslip.io]: ⏎
Pulling ghcr.io/metor-com/metor-box:latest … done
Starting box + Caddy (TLS via Let's Encrypt) … done
Your setup link (24 h, single use):
https://188-68-58-3.sslip.io/bots/auth/claim?token=…
open it on your phone

One command on a fresh Linux server.

The installer installs Docker if it is missing, pulls the image, starts metor behind Caddy with TLS, and prints your setup link as text and as a QR code. You do not need a domain: the installer proposes a name that resolves to your server's address, and a real domain can replace it any time. Already running a reverse proxy? Forward one path and keep your setup.

Sign-in without passwords.

The first browser gets in with the setup link. Every further phone or computer is linked from a device that is already in, by QR code or pairing code. A device list shows who is signed in, and any device can be removed.

Updates keep everything.

Pull and restart with compose. Bots, histories and routines survive; sessions resume with their context. Backups are three volumes.

On your Mac.

Today with Colima or Docker Desktop and a local build; a Homebrew install that sets up everything is on the roadmap.

Guardrails

Boundaries you can read, not trust us on.

  • The box is the sandbox.

    Inside its container a bot may read, write, run and browse without asking. Approvals are reserved for the boundary: your own machine, and actions with an effect outside, such as sending, paying, publishing.

  • Official login only.

    Harnesses run as processes with their own login. metor never extracts tokens or impersonates a subscription.

  • Nothing leaves silently.

    No silent uploads, no source maps in releases. Your data lives in three volumes on your machine; a backup is a copy.

  • Open source.

    Apache-2.0. The image bundles Claude Code and Codex from their official channels; using them needs your account and their terms.

  • Early.

    Version 0.1.0 works end to end, with rough edges. The changelog, the backlog and the image are public on GitHub.

Also from metor

metor memory: #1 on EnterpriseRAG-Bench.

A memory engine for company knowledge. It answers your agents with verified facts and exact counts instead of plausible guesses, and plugs into any MCP-capable harness. A product of its own.

80.34official overall score, single pass
86.2answer completeness, the highest on the board
12+systems on the official leaderboard

Move your agents in.

Five minutes on a fresh server. Your subscription, your machine, your bots.

$curl -fsSL https://www.metor.com/install.sh | sudo bash