Full access is a contract, not a checkbox
Five levels of rights, an owner ceiling checked on every call, a journal written before the answer, and a ten-minute check for any vendor.
🎯 For Server owners and teams choosing how much an AI assistant may do on their machines.
- See the five levels and who may grant each
- Understand the owner ceiling and per-call check
- Ask a vendor the five questions that matter
An AI assistant connected to your server does not run in the cloud. It asks a small program on your machine, and that program acts with the rights of the user it runs as. So the question that decides everything is not which model you chose, but which user runs the connector, which actions it accepts and who writes down what it did. This article explains how we split those rights, why "full access" is a contract rather than a checkbox, and how to check a vendor's offer in five minutes.
The question behind the question
People ask: "can the assistant fix my server?" The useful answer: it can do exactly what the connector is allowed to do, and nothing else. Everything else is marketing.
Three things decide the outcome:
- The process user. A connector running as
rootcan do everything, even if its settings screen shows a friendly list of tools. A connector running as a dedicated unprivileged user cannot, even if the model insists. - The action catalogue. A fixed list of named commands with fixed arguments is a limit. A free-form shell is not a limit, whatever the prompt says.
- The record. A journal written before the answer tells you what happened. A counter without command details does not.
Why we do not sell "full access" as one switch
In November 2026 we published the account of two days when a third-party remote-shell bridge ran as root on one of our hosts: 660 command executions and no record of which commands they were. Access was revoked, the pairing state was removed, keys were rotated. We do not claim that nothing was read — with a root process and no command log, the absence of evidence is not evidence of absence.
That is why our own connector ships with a reviewed policy file and a stop file instead of a root shell, and why "full access" in our product is a separate level with its own conditions, not a default.
The five levels
| Level | What it allows | Where it runs | Who can grant it |
|---|---|---|---|
read |
list files, read files, service status | workspace only | client |
write |
edit files inside one declared folder | workspace only | client |
exec |
build and test commands from the catalogue | connector's sandbox | client |
deploy |
deploy and publish commands | host, named commands | client, if the owner allows |
full |
commands tagged full, including shell |
host | owner only |
Two rules make this more than a menu:
- The grant is an intersection. The client chooses what it wants; the host owner writes a ceiling into the policy. The effective grant is what both allow, and the ceiling is re-checked on every call, not only when the token is issued.
- Missing scope means a missing tool. If the grant does not include
write,fs_writedoes not appear in the tool list at all. The model is not asked to behave; the possibility is absent.
What an approval looks like
For anything that is not read-only we separate three things: the request, the decision and the execution.
- The model can request: "restart service X".
- The owner decides: the platform checks the account, then signs a short-lived ticket bound to this host, this action, these exact parameters and a deadline.
- The privileged executor executes only with that ticket, applies a timeout, runs the command without a shell and writes a hash-chained journal entry before answering.
The connector edge never holds the signing key. A compromised edge can annoy you; it cannot become root.
Why the counter is part of the contract
Every call that reaches the executor is one row in the journal: command, exit code, duration, digest of the output. The monthly allowance counts those rows.
- Free tier: 10 000 calls per month.
- Above that: a paid plan (Pro, Pro Full, Lifetime). Prices are a separate offer and are not a security claim.
- A call refused by the grant is not charged; a call whose cost is unknown is
never counted as zero — it stays
uncertainuntil reconciled.
If a vendor bills "tokens in the chat" but cannot show the call log, the number cannot be checked. A counter without a journal is a guess with a currency sign.
Five-minute check before you connect anything
- Ask which user the connector runs as. "root" is an answer, not a detail.
- Ask for the list of allowed actions. If the answer is "any command", the limit does not exist.
- Ask where the journal is and what a record contains (command? exit code? digest?). If there is no journal, there is no audit.
- Ask how to revoke access to one host without touching the others.
- Ask what happens on the second call after you lower the permission. If the answer is "the old token still works", the ceiling is decoration.
What we do not promise
- We do not promise the model will never make a mistake. We promise the mistake is visible, bounded and reversible.
- We do not promise that a hidden URL is security. Access control is a grant and a revocation, not an unlisted address.
- We do not promise that "full access" is safe. It is dangerous by nature, which is why it is a separate level, time-limited and confirmed.
- We do not mix client accounts with the owner's control channel.
Where this is implemented
The privileged executor, the ticket format and the journal are in the
rootremdesk repository; the levels, the call counter and the installer are in
remdesk; the consent screen and grants live on the platform. The public policy
page and the incident record are linked from the connector's documentation. If a
claim here does not match the code you can read, the code wins.
Sources
Source links and observation dates are recorded below. Review status is stated for each material.
🔗 Article connections
Open the full knowledge graph ↗Published article links. Select a node to open its article at the heading.
Show the connection graph
Article links · keyboard / text view
Same topic
- Tvijo agent system: PRIZMA, RUNLEDGER and acceptance
- Read agent communication receipts without mistaking delivery for a human read
- Audit an AI technology radar against primary sources before promoting a story
- Run an AI coding harness evaluation offline before any live pilot
- DSHREM on a phone: SSH, tmux, and safe task handoff
- DSHREM owner practice: cwd, worktrees, WIP, and root boundaries
Next steps
Related reading
Tvijo agent system: PRIZMA, RUNLEDGER and acceptance
Tvijo is developing a controlled workflow from a repository task to an accepted result, with permissions, resource limits and attributable evidence. PRIZMA describes the intended resource profile; RUNLEDGER connects the work to its checks a…
AI and process foundationRead agent communication receipts without mistaking delivery for a human read
Find the current stage of a swarm message without confusing durable storage, delivery, runtime receipt and viewing. This guide connects local checks, the private observation screen and distinct receipts, so a delivery acknowledgement does not become an unsupported claim that an action completed.
AI and process foundationAudit an AI technology radar against primary sources before promoting a story
Check a technology news item against its primary source, version and event date before deciding to adopt it. This guide follows a fixed ten-item sample, separates facts from interpretation and forecasts, and preserves limitations and the condition for checking the claim again.
💬 Comments