Security

Safe isn't a promise.
It's plumbing.

We don't ask you to trust the agent. We built the pipe so that a confused, compromised or simply over-eager agent has nothing to steal, nowhere to wander, and no way back in once you revoke it. Here is the whole model, beat by beat.

One command's life,
start to finish

Every command an agent sends lives this exact loop — there is no other path to your servers:

  • The policy gate sees it first. Allowed, denied, or parked for human approval — before anything touches the network.
  • The relay mints a certificate. Signed by your org's own CA, valid for five minutes, pinned to our relay's egress address. The agent never sees it.
  • A sandbox opens to one server. The session runs in isolation and can reach exactly the machine it opened against — nothing else.
  • The internet gets nothing. Your firewall allowlists the relay IPs (or your dedicated IP); direct attempts on port 22 die at the ring.
  • The certificate dies on its own. No rotation calendar, no cleanup. Five minutes later it's a brick.

Four principles,
no exceptions

Agents hold nothing

Your SSH keys are sealed in the vault with your org's encryption key and never leave it. Agents authenticate to the portal with a revocable token; servers see only short-lived certificates minted per session. There is no credential on the agent's side worth stealing — a leaked token is useless for SSH and dead the moment you revoke it.

Blast radius of one

Scope caps what an agent can see; the session sandbox caps what a session can reach — exactly one server, the one it opened against. A prompt-injected agent can't pivot to the box next door, because the plumbing between them doesn't exist. Scope is re-checked on every call, so detaching a server takes effect mid-conversation.

Evidence, not logs

Sessions are recorded into a hash-chained audit trail: every entry carries the hash of the one before it, so a tampered record breaks the chain visibly. Commands, outputs, file transfers, approvals and denials all land in it. Replay any session after the fact and settle "what exactly ran?" with footage.

Revocation is total

Click revoke and the agent's token dies, its live sessions are torn down mid-command, and there is nothing to clean up afterwards — it never held a key, and its certificates expire within minutes regardless. Same for a departing teammate. Nothing to rotate, nothing to doubt.

The boring details,
on purpose

  • Per-org certificate authority. Each org gets its own CA; you install one TrustedUserCAKeys line per server. Roll the CA any time and every outstanding certificate dies with it.
  • Certificates pinned to source. Issued certs carry a source-address critical option — they only work from our relay's egress. Exfiltrated mid-flight, they're still useless.
  • Browser connectors use OAuth. claude.ai, ChatGPT and friends get a consent page and their own revocable grant — your account password never travels.
  • Hosted agent runs are caged. CLI runs execute in locked-down containers whose only network egress is the portal itself. No side channels to the internet.
  • Private ranges are off-limits. Customer targets are checked against reserved and private network blocklists, every time, so the portal can't be aimed at infrastructure that isn't yours.
  • Humans follow the same rules. The browser terminal rides the identical relay, policy, certificate and recording path as the agents. No privileged side door for staff — yours or ours.
your-server · /etc/ssh/sshd_config
# 1. trust your org's CA (one line, one time)
TrustedUserCAKeys /etc/ssh/sshportal_ca.pub

# 2. optional: let only the portal reach port 22
#    ufw: relay IPs or your dedicated IP only
ufw allow from 198.51.100.7 to any port 22
ufw deny  22

# that's the whole server-side footprint:
# no agent daemon, no outbound tunnel, no root helper
works with stock OpenSSH nothing to install on your servers

Give the agent hands.
Keep every key.

One CA line on your servers, one MCP URL in your agent. The model above does the rest.

Get started · $9.95/mo