remoteagent.online · open source
Every dream starts as a sentence. Give it a real machine, your real files, and a record that cannot be rewritten afterwards — and it stops being a maybe.
The cloud queues what it may do. Your device decides what actually happens.
$ curl -fsSL https://remoteagent.online/install.sh | bash
$ remote-agent login # connect this device with your Remote Agent key
$ remote-agent gui # or: remote-agent start (headless)
A Node daemon (remote-agent) on your machine, plus the cloud console. Your files never move: the cloud queues typed operations and your device executes them.
The device client is public at github.com/airemoteagent/AI-Remote-Agent; the server side is the mona-plugin (1.2.0) in this app. You can read both.
One LLM provider key — OpenAI, Anthropic, Google, DeepSeek or OpenRouter — and a machine to run the agent on. The key is stored encrypted on the server, never on the device.
In plain words
How it works
One pipeline for every action. Nothing skips the gate in the middle.
The daemon polls for work and claims each task atomically, so only one of your machines executes it. The console shows which device is online and what it is running.
Workspace sync sends labels, counts and an identity hash — no host paths, no file bodies. Read, write, patch, rename, copy, trash and diff all run locally on the device.
Rules are first-match and deny by default. On POSIX, shell commands are argv-only with an allow-list and a scrubbed environment; risky patterns can require your approval; requests are rate-limited.
Outbound requests resolve DNS first and refuse loopback, private, link-local, metadata and CGNAT ranges. Redirects are validated again, with 50 KB and 15-second caps.
A run moves created → running → verifying → succeeded / failed / cancelled with per-step checkpoints. A crash resumes instead of repeating, and side-effecting retries require an idempotency key.
Tool output, files and web pages are labelled untrusted in the agent's own prompt and are never followed as commands.
Plan limits are checked server-side: agents, runs per day, steps per task and requests per minute. Free is 10 agents · 200 runs/day · 8 steps · 60/min; Pro raises all four.
Provider keys are stored AES-256-GCM encrypted. The API returns them masked, never in full, and sensitive actions ask for a second factor first.
The device keeps a hash-chained local audit log and anchors its chain head to the cloud (append-only: sequence, hash, signature). The console keeps the per-step trace of every run.
The console
Click through it — this is the real console, panel by panel.
Get your first agent running in 3 steps
The real dashboard keeps these live: agent status, device telemetry, runs and spend.
Real files. Visible changes. Verified delivery.
<!doctype html> <html lang="en"> <head> <meta charset="utf-8"> <title>My app</title> </head> <body> … </body> </html>
Renders, no console errors
Verdict recorded against the run
Your files stay on your machine — the cloud only queues the operations your device runs.
Your helpers and their rules
20 built-in tools, each one yours to switch off
Receipts: steps, tokens, cost
Example 1 from the published docs — a real captured run.
Brain mode for this run — the console compares them from real runs:
What it cost — per day, model and agent
The three published one-prompt builds — real captured cost, drawn to scale.
What waits for your OK
Approve or deny — the decision is recorded with the request, and denying asks for a note.
The console's own empty state: “No pending approvals — everything is settled…”
Skills and the command allow-list
6 built-in skills — written procedures the agent follows, switched on or off per agent.
232 commands in 9 real categories — the allow-list you pick from, or allow everything with *.
Custom commands (dynamic tools) you add are pushed to your devices automatically.
Everything else, organised
Plan limits are enforced server-side. A user-set spending cap is not built yet — the console says so here.
Click any section. Layouts are the real console; anything illustrative is labelled, and every figure shown is real.
File tree and editor, diff, sandboxed preview with start/stop, versioned changes with content hashes, acceptance criteria, evidence with a verdict, agents granted review-only or read/write access, and single-use delivery links.
Every run keeps its steps, tokens and cost. Watch a live trace, read the audit trail, filter the live event feed (all · chat · tools · brain steps · errors) and export runs as JSONL for training.
Tasks, brain calls, tokens and cost — per day, provider, model and agent. Brain modes are compared from real runs: auto · cheap · balanced · deep · manual.
Per-agent model and system prompt, capabilities, skills and tool toggles (20 built-in tools), shell allow-list, allowed file paths, per-agent chat, start / stop / update.
Live telemetry: CPU, memory, disk, uptime, load, cores, version and last seen. One Remote Agent key connects a machine, and device tokens can be revoked.
Skills synced from ~/.remoteagent/skills, plus 232 commands across 9 categories: pick an allow-list or allow everything with *. Your own commands (dynamic tools) push to your devices.
Test any model with a stored key and see the response, provider, token usage and latency before an agent ever uses it.
Two-factor authentication (authenticator, backup codes, email code), approval requests you approve or deny, CSRF-protected writes, and every query scoped to your user.
Cost is measured and reported per run. A user-set spending cap does not exist yet — the console says so on the settings page. Plans limit agents, runs and steps, not euros.
Chat
Every reply comes back with the steps behind it — what it thought, which tool it called, what came out, and what it cost.
Example 1 from the published docs — a real captured run: 9,947 tokens · ≈ $0.0028 · deepseek-chat. The layout is the real chat hub.
The live trace streams while a run is going: think, tool.call, tool.result, answer, verify. If nothing is recorded yet, the console says exactly that instead of filling the space.
Cancel a running task, thumbs-up or thumbs-down a reply, and pick which workspace an agent may touch before you send.
Your first message creates the helper automatically. If no device is online yet, the cloud answers inline instead of failing.
Workspace
Your files stay on your machine — the cloud only queues the operations your device runs.
<!doctype html> <html lang="en"> <head> <meta charset="utf-8"> <title>My app</title> </head> <body> … </body> </html>
Renders, no console errors
Verdict recorded against the run
UI preview — the real workspace layout, with example content.
Read, write, patch, rename, copy, trash, restore, diff. The cloud queues typed operations; the device executes them with a lease and reports the result hash back.
Every save becomes a version with a before → after hash, so you can see what moved instead of trusting a summary.
Start and stop a preview server, opened in a locked-down frame — default-src none, nothing else reachable from it.
Write the criteria, let a run produce the evidence, and keep the verdict with the workspace. When a device mutates a file, the run is verified against the recorded hash and conflicts are reported.
Grant an agent review-only or read/write access, and mint a single-use delivery link when the work is done.
Examples
Traces published in the docs, captured by the control plane and anonymized — nothing re-drawn.
And 232 commands across 9 categories — the allow-list you pick from, or allow everything with *.
Six built-in skills ship with the agent: briefing, disk-health, network-health, service-health, web-research and certificate-expiry. A skill is a written procedure the agent follows, and you can switch each one on or off per agent.
Add a dynamic tool (name, description, command, parameters) and it is pushed to your devices automatically — no redeploy.
Download finished runs as JSONL with steps, tokens, cost and your thumbs-up or thumbs-down — training data from your own work.
For you
Same agent, different jobs. Every card names the real tools or the real skill behind it.
It runs your dev commands, reads the failures, and explains them in order of importance.
Edits land in a workspace with a diff, a content hash and a version you can roll back to.
Searches, renames, copies — and moves deletions to a trash you can restore from.
One message: how the machine is doing, the top headlines, and anything you asked it to remember.
Checks every volume, flags the ones over 85%, and proposes one safe cleanup — it never deletes on its own.
Launch apps, open a page, speak a line out loud, or convert a photo — the everyday commands, on your side.
Two or three different searches, the best pages fetched, disagreements called out, a note saved for next time.
Walks the tree, reads the files, and answers in plain language. Ask again tomorrow — it remembers.
Reads your own notes and documents, then quizzes you on them. Your material never leaves your machine.
Collects status and safe recent logs, proposes a restart, and waits for your OK. It does not restart on its own.
Runs read-only checks, classifies the fault, prepares a plan — and changes nothing without approval.
Reads validity dates, flags what is near expiry, prepares the renewal. Private keys are never read or shown.
Skill names and commands come from the real catalogue. The runbooks diagnose and ask; they do not act alone.
Install — a real app, not a website
Install it once. After that you just type what you want — in the browser console, or in the app on your machine.
irm https://remoteagent.online/install.ps1 | iexRun it in PowerShell. It installs the remote-agent command onto your PATH.
curl -fsSL https://remoteagent.online/install.sh | bashApple Silicon and Intel. Then remote-agent login once.
curl -fsSL https://remoteagent.online/install.sh | bashSame installer. Works behind NAT — there are no ports to open.
curl -fsSL https://remoteagent.online/install.sh | bashARM boards included. Leave it on as your always-on helper.
remote-agent gui.Your device dials out to the cloud. Nothing connects in, so nothing needs opening on your router.
Platforms, licence and requirements are the real published ones.
remoteagent.online × llminone.com
Two open-source projects, two halves of one wish: one makes it real, the other keeps it alive. The thing you imagine in the browser does not die as a demo — it gets a machine, a budget and a memory.
① Create — llminone.com
One sentence becomes a working thing, live in your browser.
llminone.com ↗
② Keep alive — remoteagent.online
The build gets a home: real files, a keeper and a record.
remoteagent.online ↗
It does not end when the demo ends. One vault. One budget. One record — every dream gets a keeper.
not integrated yet — the direction we are building toward