A real terminal
Full xterm in the browser, keyless, on any server in the fleet. No VPN, no bastion hops, nothing to install on your laptop.
Platform
sshportal.ai sits between anything that speaks MCP and every server you own. Agents get nine purpose-built tools instead of a raw socket; you get a policy gate on every command, an approval queue for the risky ones, and a replayable record of all of it. Humans share the same fleet through a real browser terminal — same rules, same audit trail.
Claude Code, claude.ai, Cursor, Codex, Gemini, Grok, your own tool loop — every client points at https://sshportal.ai/mcp. OAuth consent for browser connectors, a bearer token for CLIs. That's the entire integration.
Scope decides which servers an agent can even see. Policy decides what runs, what's blocked, and what waits for a human. Both are re-checked on every single call — not once at connect time.
Person or agent, every session lands in a hash-chained, tamper-evident audit log with full replay. "What did the agent do at 3 a.m.?" is a video, not a guess.
Raw SSH gives an agent a socket and wishes it luck. The portal gives it a vocabulary that matches how agents actually work: discover what's in scope, run one-shot commands, or hold a full interactive session — install wizards, REPLs, htop, anything with a PTY.
Sessions are first-class: an agent can open one, keep state between steps, and your teammates can attach to the same PTY from the console and watch — or take over.
list_endpoints only ever returns the servers you attached to that agent.run_command for the quick read, sessions for the stateful fix.wait_for_approval lets the agent block politely until you click.list_endpoints servers this agent may touch
run_command one command, one server, output back
open_session interactive PTY on an endpoint
send_input type into a live session
read_output read what the session printed
list_sessions this agent's live sessions
attach_session rejoin one that's still running
close_session tear it down cleanly
wait_for_approval block until a human decides
Rules are yours: regex or prefix patterns, org-wide or per-server, each ending in allow, deny, or require approval. Explicit rules always run first; the org-wide mode decides what happens to everything else.
rm -rf, schema drops) and send them to the queue; the rest runs.Approvals land in the console and the agent waits in-band. One click and it continues mid-conversation; deny it and the agent gets told no, with the reason on record.
allow ^(df|du|free|uptime|systemctl status)
allow ^journalctl read the logs, always
require_approval ^systemctl (restart|stop)
require_approval ^(reboot|shutdown)
require_approval ^apt(-get)? (upgrade|dist-upgrade) prod only
deny rm -rf /
deny ^DROP (TABLE|DATABASE)
[queue] ops-bot on db-01: systemctl restart postgresql
[queue] ▸ Approve · Deny
The portal is more than the pipe. The console is a full operations room for the same fleet — for the humans on the team and the agents beside them.
Full xterm in the browser, keyless, on any server in the fleet. No VPN, no bastion hops, nothing to install on your laptop.
Drag a file onto the terminal and it lands on the server; pull one back just as easily. Every transfer is logged with who, what and where.
Open any live session read-only and watch the agent work, keystroke by keystroke — or attach to the same PTY and drive beside it.
Sessions are recorded into a hash-chained log. Scrub through any session after the fact and prove to anyone exactly what ran.
Launch Claude Code, Codex, Gemini or Grok CLIs in a locked-down container we host. Its only network reach is your fleet, through the same scoped MCP tools.
No agent of your own yet? Chat with your fleet straight from the console. Same scopes, same policy, same recordings — batteries included.
Reach your servers from an address that's yours alone. Allowlist it in your firewall and port 22 goes dark to the rest of the internet.
Invite the whole team — seats are free. Members get scoped access and the same keyless treatment; owners manage policy, approvals and billing.
Add a server, deploy an agent, paste one MCP URL into your tool of choice.
https://sshportal.ai/mcp