Self-hosted Chrome sessions for you and your AI agents. MCP, live human takeover, saved profiles
Self-hosted Chrome sessions for you and your AI agents. MCP, live human takeover, saved profiles, drag-and-drop files, downloads and audio.
·
Truly secure browser sessions for you and your agents.
Give every agent its own real Chrome on your own server. Watch it live, and take over with one click when it hits a login, a 2FA code or a CAPTCHA. Then hand it back.
▶ Try the live demo · Features · Documentation · Quick start · Compare
English · Português (Brasil) · Español · 简体中文 · हिन्दी · Français · Deutsch · 日本語 · Русский · Bahasa Indonesia · 한국어
Agents that use the web need a browser. Most teams rent one by the hour from a cloud service, or glue together a headless Chrome and a VNC server. easi.run is a third way: one command starts it on your own Linux server, and from then on every agent gets an isolated Google Chrome, fresh or with your saved logins, and you get a live window into each one.
When the agent gets stuck at something only a person can do, you take control from the session's bar. The agent waits. You type the password, approve the 2FA prompt, or solve the puzzle. You give control back, and the agent carries on.
[redacted]; your own chat keeps seeing the screen (how).easi.| easi.run | Browserbase | Browserless | KasmVNC | |
|---|---|---|---|---|
| Runs on your own server | ✅ | – | ✅ | ✅ |
| No usage fees | ✅ | – | – | ✅ |
| No plan limit on browsers at once | ✅ | – | – | ✅ |
| No plan limit on session length | ✅ | – | – | ✅ |
| Take over, and the agent waits | ✅ | – | ✅ | – |
| Your agent calls you in, and you’re told | ✅ | – | ✅ | – |
| Build your own takeover screen | ✅ | – | – | – |
| An API your agents can call | ✅ | ✅ | ✅ | – |
| An MCP server for your agents | ✅ | ✅ | ✅ | – |
| Your typing hidden from the agent | ✅ | – | – | – |
| Encrypted session vaults only your password unlocks | ✅ | – | – | – |
| Private network off-limits | ✅ | – | – | – |
| Built-in stealth | ✅ | ✅ | ✅ | – |
| Live sound | ✅ | – | – | – |
| Drag a file straight onto the page | ✅ | – | – | – |
| Your own file picker for uploads | ✅ | – | – | – |
| Downloads land on your computer | ✅ | – | – | – |
| Copy and paste both ways | ✅ | – | ✅ | ✅ |
| Record the session | ✅ | ✅ | ✅ | – |
| A timeline of the agent’s actions, with a replay | ✅ | ✅ | – | – |
| Your own Claude Code or Codex beside the live browser | ✅ | – | – | – |
| Turn off input, clipboard, files or sound per session | ✅ | – | – | ✅ |
| Your own proxy per session, location matched | ✅ | – | – | – |
| Works over one HTTPS port, no TURN | ✅ | – | – | ✅ |
| Open source you can audit | ✅ | – | – | ✅ |
✅ has it · – doesn't, partly, not in their docs, or not applicable. The full comparison covers 19 alternatives, with the sources and notes (October to November 2026).
| Minimum | Comfortable | |
|---|---|---|
| Server | Any Linux x86-64 machine with Docker: a VPS from any cloud (Hetzner, DigitalOcean, AWS EC2 x86…) or your own server | |
| Software | Docker Engine 28+ (rootful, run by docker.service) with Compose and Buildx, systemd, git, curl, flock, iptables, ss, awk, stat, df, timeout and realpath | |
| Account | A non-root user in the docker group (sudo usermod -aG docker $USER, then log in again); no sudo needed: the installer makes its few changes as root through a short-lived privileged Docker container, which Docker must allow | |
| CPU | 2 cores | 4+ cores |
| Memory | 4 GB (one browser session at a time) | 8 GB+ (2 sessions at once on 8 GB, 4 on 16 GB) |
| Disk | 10 GB free (the images), and what replays take: up to 10% of the disk, set in Settings | |
| GPU | none |
Each browser session runs with its own limits: 2 CPUs and, for Chrome, 2 GiB of memory or what your server's memory leaves each session, up to 8 GiB, plus a display, proxy and agent
beside it (1152 MiB together); an open chat adds up to 2 GiB and its proxy 128 MiB. The installer runs as many sessions at once as
your server's memory holds, from 1 to 16, and its summary says how many and how much memory each browser gets; each starts when you or an agent asks for
one. To change these, put BROWSER_CPUS, BROWSER_MEMORY_MIB, CHAT_MEMORY_MIB or ADMIN_MAX_SESSIONS in ~/easi.run/.env, then run
bash install.sh again (Configuration).
Ask your coding agent (Claude Code, Codex, or another) on your server:
Read https://easi.run/install.md and install easi.run on this server.Or run it yourself:
curl -fsSLO https://easi.run/install.sh && less install.sh && bash install.shIt checks the server, puts the firewall rule in place that keeps pages off your server's own services, sets the
server's TCP for viewers far away (fq and BBR, when the kernel has BBR), starts easi.run, and adds it to Claude Code
when Claude Code is installed. It needs no sudo: the firewall rule and the TCP settings are made as root through a
short-lived privileged container of Debian's image, pulled by its digest, whose nsenter runs your server's own
iptables, systemctl and sysctl. Its summary tells you what it did, and where you create the admin's password.
Nothing is compiled on your server: the installer pulls the images easi.run's CI built and published, each checked
against the digest the installer names, and adds Google Chrome to the browser image from
Google's signed repository (Google's terms let no one else ship Chrome). The first run downloads about 550 MB and
takes one to two minutes on 2 cores with a fast connection.
On the server: https://localhost:9443.
From your computer: use the command you already use to log in to the server, adding
-N -L 9443:127.0.0.1:9443, and keep it running; then open https://localhost:9443. For example:
ssh -N -L 9443:127.0.0.1:9443 you@your-serverThe installer's summary prints this line, filled in with your login when it can tell it.
VS Code or Cursor over Remote-SSH forward the port for you: open the link from their terminal, or add port 9443 in their Ports view; either way, the local port must be 9443 too.
The certificate is your server's own, so the browser warns you about it: the installer keeps a copy in
~/.config/easi-run/tls-cert.pem, and the admin logs its fingerprint (docker compose logs admin | grep fingerprint,
in ~/easi.run) for you to compare.
Open https://localhost:9443 (from your computer, through the SSH tunnel of step 2) right after the install: until a
password is set, the admin's first screen asks for one. Choose it, 12 characters or more; you are logged in at once.
Next, the admin offers a second step at sign-in. It is optional, and we suggest it: with it on, your password alone opens nothing.
Turn either on, off or replace it later under Security. Lost your phone or passkey and your recovery codes? Run
docker exec easi-run-admin-1 admin reset-second-step on the server, and the password alone signs you in again.
From then on, log in, start a session, and take control from its bar whenever you want.
curl -fsSLO https://easi.run/install.sh && bash install.sh --updateIt moves easi.run to the release the newest installer names, and removes the images of the one before; sessions running at the time may end.
bash ~/easi.run/install.sh --uninstallAfter you confirm, it removes easi.run's containers, volumes and images, its firewall rule and the docker.service
drop-in that puts it in place, and its TCP settings, putting back the values they replaced (unless another easi.run
installation on the server still needs them),
its own files in ~/.config/easi-run (the directory too, when nothing else is in it), and Claude Code's easi server
when it is this installation's.
It keeps the checkout, ~/easi.run, Docker's build cache, and the Debian image its root steps ran in, and says how to
remove them.
Without an SSH tunnel, you can put one of these in front of the admin, each with your own account. Exposing the admin
beyond the server is your choice. Either way, set ADMIN_ORIGIN to the address you will open
(Operations has the steps):
sudo tailscale serve --bg https+insecure://localhost:9443 serves it at
https://<machine>.<tailnet>.ts.net, to your tailnet only; the admin stays on the server's 127.0.0.1.cloudflared carries it to a name on your domain, and Access checks
who you are before any request reaches the admin's login page.What protects it then: who may reach the address at all (your tailnet's devices, or your Access policy), then the admin's sign-in: your password and, when you turned it on, the second step, or a passkey.
Every session your agent starts comes with a watch_url: open it to watch live and take over.
When the installer's summary says MCP is ready, easi.run is in Claude Code and Codex (each one installed for the account that ran the installer): start a new session and ask it to browse. For another account, or any MCP client that runs a command:
codex mcp add easi -- docker exec -i easi-run-admin-1 admin mcpThe same command, carried by SSH (log in with your SSH key): no token, port or certificate to set up.
claude mcp add --scope user easi -- ssh -T you@your-server docker exec -i easi-run-admin-1 admin mcp
codex mcp add easi -- ssh -T you@your-server docker exec -i easi-run-admin-1 admin mcpUse the login you already use for the server in place of you@your-server. Keep the SSH tunnel from step 2 open to
watch the sessions in your browser.
Create a token on the admin's Agents page, and use MCP at /mcp or the same tools under /api/agent/
(Agents guide).
Start a session with CDP for Playwright, Puppeteer and Selenium on: its bar shows the port (CDP :9222; a click
copies http://127.0.0.1:9222), on the server's loopback. Attach to the browser that is already there:
const browser = await chromium.connectOverCDP('http://127.0.0.1:9222');
const page = browser.contexts()[0].pages()[0];Puppeteer takes puppeteer.connect({ browserURL: 'http://127.0.0.1:9222' }), and Selenium Chrome options'
debuggerAddress 127.0.0.1:9222. From your computer, bring the port over SSH first:
ssh -N -L 9222:127.0.0.1:9222 you@your-server. Whatever attaches drives the whole browser and reads what it holds,
what you type in it included, and anyone on the server can attach: turn it on for the sessions you script.
Start a Persistent session, log in to your sites in it, and stop it: it keeps, encrypted, the logins a site keeps
when Chrome closes, and brings them back when it is resumed. Allow the agent's token (or the local MCP clients) on that saved session; the agent then finds it with
list_saved_sessions and resumes it with start_session's resume.
Against KasmVNC on the same host and CPU budget, medians of the headless runs (2026-10-21, over the WebSocket):
| easi.run | KasmVNC | |
|---|---|---|
| Smooth scroll | 55 fps, 1.8 Mbit/s | 24 fps, 12.4 Mbit/s |
| Full-motion animation | 38 fps, 2.0 Mbit/s | 24 fps, 25.4 Mbit/s |
| 1080p video | 36 fps, 2.1 Mbit/s | 24 fps, 18.1 Mbit/s |
| Scrolling text | 1.9 Mbit/s, SSIM 0.990 | 16.4 Mbit/s, SSIM 0.977 |
| Typing, key press to drawn frame | 28 ms | 69 ms |
| Static page | no media traffic | no media traffic |
Every metric, the method and the raw results are in the benchmark.
flowchart LR
person(["Your browser"])
internet(("Internet"))
subgraph session["One session per browser"]
agent["agent<br/>Cua Driver"]
browser["browser<br/>Google Chrome"]
display["display<br/>Rust compositor"]
proxy["proxy"]
end
admin["admin<br/>HTTPS, and UDP for the media"]
controller["controller"]
docker[("Docker")]
browser -- Wayland --> display
agent -- input --> display
browser -- Unix socket --> proxy --> internet
display -- tiles, video, sound --> admin
admin <-- HTTPS and WebSocket, WebTransport over UDP --> person
admin --> controller --> dockerA small compositor written in Rust (display) draws Chrome's windows into its own memory. Still content travels as
lossless tiles; motion stays lossless while that is cheap, and otherwise streams as AV1 video that sharpens to
pixel-perfect once motion stops; the page's sound travels as Opus, over one HTTPS port, and over UDP
(WebTransport) where your network lets it through, which keeps the stream smooth on lossy networks. Only the viewer
who holds control can type and click.
Issues and pull requests are welcome. CONTRIBUTING.md covers building, testing and review, and
scripts/dev/run.sh scripts/check.sh runs the checks a change needs in the same image CI uses. Please report security
problems privately, as SECURITY.md describes.
easi.run is free software under the GNU Affero General Public License, version 3 only. Run it, change it,
and share it; if you offer a modified version to people over a network, offer them its source too. Every dependency
keeps its own license, listed in docs/licenses.md. The published images encode video only as
AV1, which is royalty-free; an H.264 encoder is an opt-in build feature (--build-arg DISPLAY_FEATURES=h264). They
are published as ghcr.io/julianosig/easi.run-<role>, each tagged and labelled with the commit of this repository it
was built from, which is its source; the browser image carries no Google Chrome, which each installation adds itself.
Each carries, in /usr/share/doc/easi.run, this license, where its source is, and the license and notice texts of what
it ships.
If easi.run saves your agents a headache, a ⭐ helps others find it.
Paste this prompt into your agent. It reads this page and does the setup for you.
Read https://aiagentslisting.com/mcp/easi-run to learn what the "easi.run" MCP server does and how to install it. Add it to my coding agent's MCP configuration as documented on that page, then confirm the server connects and list the tools it exposes.Agents can also browse this directory over MCP at https://aiagentslisting.com/api/mcp
claude mcp add --scope user easi -- ssh -T you@your-server docker exec -i easi-run-admin-1 admin mcpThis server doesn't publish a machine-readable tool manifest, and it doesn't expose a public endpoint we could ask. Connect the server locally to see what it exposes.
We could not run this server: it ships no npx package for our sandbox and no public endpoint. The four checks that need a live server are left out, and the grade tops out at B until one exists.
Paste this line near the top of your repository's README. The badge always renders the current score.
[](https://aiagentslisting.com/mcp/easi-run)BlockVectra MCP for AI agents: Robinhood Chain and multi-chain EVM RPC; keyless docs/discovery; RPC
MCP server for B2B Social Video Transcripts: transcripts and outbound signals from TikTok, Reels, X.
Find the Reddit threads that name a company, its competitors.
No tagged releases on record.
Connect to the AI Agents Listing MCP server at https://aiagentslisting.com/api/mcp using the HTTP transport and complete the sign-in when it asks. Then call the submit_review tool for the listing "easi-run" with a rating from 1 to 5 and a short note on what worked and what did not.No reviews yet
Be the first to share your experience with this tool.
BlockVectra MCP for AI agents: Robinhood Chain and multi-chain EVM RPC; keyless docs/discovery; RPC
Developer Tools
MCP server for B2B Social Video Transcripts: transcripts and outbound signals from TikTok, Reels, X.
Browser & Scraping
Find the Reddit threads that name a company, its competitors.
Browser & Scraping
One email a week. New agents, MCP servers and skills, and what is actually getting traction.