Developing from your phone with an AI chat: cloud first, deploy later
Let the assistant work in a cloud repository from your phone and deploy yourself from a device that holds the key. Why a phone connection with deploy rights is a door through the chat provider.
🎯 For Developers and server owners who code from a phone through an AI chat connected to their infrastructure.
- Understand that a connector token lives with the chat provider
- Split work into a cloud zone and a production zone
- Choose remdesk boxes for a phone connection
- Deploy from a device that holds the key
Use an AI chat on your phone to write code in a cloud repository, not to change your server. Let the assistant work on a branch, read the diff, merge on the phone if you like, and deploy later as a separate human step: from a laptop with full access or from an SSH terminal where you hold the key. If you connect the chat straight to your production server with deploy or root rights, you open a door that also passes through the chat provider, and a breach on their side becomes a breach on yours.
Why a phone plus a connector is a special case
A connector such as remdesk is a small program on your server that an AI chat can call. When you add it, you do not log in from the phone yourself. The chat provider's servers receive a token for your connector and call it on your behalf.
On our own connector, mcp.tvijo.com, sign-in is OAuth with PKCE. The access token lives 15 minutes, and the chat app renews it with a refresh token in the background. So for as long as the connection exists, a working key to your server sits with the chat provider, not on your phone.
That has three consequences:
- The provider is in the path. If the provider's systems or its connector tokens are compromised, the attacker can call your connector with exactly the rights you granted. You cannot rotate a key you do not hold; you can only revoke it.
- The phone is in the path. A phone is lost, borrowed or unlocked in a café more often than a laptop. A signed-in chat app on it can start actions on your server.
- The model is in the path. The model reads files and web pages. Text planted in a repository or a page can steer which allowed action it picks next.
None of this is a reason to stop using AI on the phone. It is a reason to decide what the phone is allowed to reach.
The safe pattern: cloud first, deploy later
Split the work into two zones with a clear handover between them.
Zone 1. Cloud workspace (phone and chat welcome). A code host such as GitHub, GitLab or a self-hosted Gitea, or an isolated workspace folder that holds no secrets. The assistant creates a branch, edits files, runs tests and opens a pull request. Mistakes here cost a revert.
Zone 2. Production (human step only). Your server, its services, its keys and its data. Changes arrive here only as a reviewed merge, deployed by you from a device that has full access: a laptop with your SSH key, or an SSH terminal app on the phone where the key never leaves the device.
The handover is the pull request. The assistant can prepare everything up to it; it does not cross it.
| Step | Where | Who | What a breach at the chat provider gives |
|---|---|---|---|
| Write code, run tests | cloud repo or workspace | assistant | access to that repo or workspace |
| Review diff | phone, code host app | you | nothing new |
| Merge | code host | you | nothing new, if merging needs your account |
| Deploy | laptop or SSH terminal | you | nothing, the chat has no deploy rights |
If you skip zone 1 and let the chat run deploy commands on production, the right-hand column becomes "your production server".
What remdesk tiers mean for this
remdesk shows the same five checkboxes on every surface that connects to it: read, write, exec, deploy, full. A grant is what you tick, limited by the ceiling the host owner set; the ceiling is checked on every call. The prices below come from the draft permission catalogue of 4 November 2026. They are planned, not on sale yet.
| Plan | Price (planned) | Boxes | Good fit for a phone? |
|---|---|---|---|
| Free | free, 10 000 calls a month | read, write, exec inside the workspace |
Yes. This is zone 1. |
| Pro | $20 a month in TvijoTokens | adds deploy, no call limit |
Only from a trusted device. deploy is never ticked in advance, and publishing needs your approval of the exact payload hash. |
| Pro Full | $35 a month in TvijoTokens | adds full: commands the owner tagged full, including free-form ones |
No. Keep it for a laptop session you are watching. |
| Lifetime | $255 once, remdesk on your own server, installation included | all tiers; you are the host owner and set the ceiling | Set the ceiling so that phone connections stay at Free boxes. |
None of these tiers is root. Root actions belong to a separate executor, remdeskroot, which is designed around owner-signed one-time tickets and is not deployed.
What each box risks, in the catalogue's own words:
read: everything inside the allowed roots goes into the assistant's context and to its model provider. A key left in a file inside the roots leaks.write: the assistant can overwrite code that someone runs later.exec: builds and tests run real code from the repository.deploy: changes reach production and other people; some effects cannot be taken back.full: whatever the owner taggedfull. You own the consequences.
A setup we recommend
- Two connections, not one. Connect the chat on your phone with
read,writeandexeconly. If you needdeployat all, grant it on a separate connection that you use from a laptop, and revoke it when you are done. - Workspace roots hold no secrets. The
.envfile, SSH keys, exchange and provider keys live outside the folders the connector can read. - Production is not a workspace. Point the connector at a checkout or a staging copy, never at the directory your live service runs from.
- Merging needs you. Protect the main branch on the code host so that the assistant can open a pull request but cannot merge it.
- Deploy from a device that holds the key. A laptop, or an SSH terminal app on the phone with a key generated on that phone and a passphrase. The chat never sees that key.
- Know your stop switch. On remdesk it is
touch /etc/remdesk/STOP: every call is refused on the next request. An access token already issued lives up to 15 minutes; only the stop file refuses instantly.
Our own terminal for the phone browser: tvijoterm, not deployed yet
People ask for a terminal they can open in a phone browser by address, without a third-party app, installed once on their server. We are building one: tvijoterm, a real root shell in the browser at https://<ip>:<port> (vim, htop, tmux and journalctl -f work). It is not deployed yet. It is different from the FinEdg developer terminal (SPEC-159), which edits project files through reviewed diffs and deliberately has no shell.
A root shell in a browser is the most powerful access path in this article, so its design starts from "closed":
- The port is closed by default. The service runs, but nothing listens until the owner opens a time-limited window on the server:
tvijoterm open 30m, ortvijoterm pairfor 10 minutes. When the window ends ortvijoterm closeruns, the port is released and every session is hung up. - A new device pairs only from the server console.
tvijoterm pairprints a one-time link valid for 10 minutes. Pairing completes only with a correct code from an authenticator app (TOTP); later devices never see the secret. - Every login needs a code. A device cookie alone does not open a session. Five wrong codes in 15 minutes lock login for everyone for 15 minutes.
- TLS always. Without your own certificate it generates one for the server's IP and prints its fingerprint; compare it once on the phone. With a domain, use a normal certificate and the warning disappears.
- A hash-chained journal records pairing, logins, failures, window open and close, sessions and revocations, but not what you typed. Revoking a device ends its sessions within seconds.
What the AI may do with it. Installation goes as an uploaded file plus an action the owner approved with a ticket, not as a download from a public link. After that, through remdeskroot, the AI may only open a window (with a ticket) or close one. There is deliberately no pair action: if the model could print a pairing link, a compromised chat account would mean a new device with root. Pairing stays a human at the console.
What stays a risk. It is still a root shell: a compromised phone with an open session is root. That is why windows are short and every login needs a code. A self-signed certificate protects you only if you compared the fingerprint; on the first pairing without that check, someone in the middle could show you their own page.
Until tvijoterm ships, an SSH terminal app with a key that lives only on the phone is the honest option for zone 2.
Check your setup in five minutes
- Which boxes does the connection on your phone have? Open the connection list on
platform.tvijo.com/connectionsor in the chat app's connector settings. - Does any phone connection have
deployorfull? If yes, revoke it and reconnect with fewer boxes. - Can the connector read a folder that contains
.env, keys or a live database? If yes, move them out of the roots. - Can the assistant merge into the main branch on its own? If yes, turn on branch protection.
- Do you know how to deploy without the chat? If no, that is the first thing to set up.
⚠️ When this does not work
The pattern assumes that deploy is a separate step. If your service deploys automatically on every merge, then merging from the phone is deploying from the phone. Add a manual approval to the pipeline, or treat merge as zone 2.
A cloud repository is not automatically safe. If it holds secrets or a pipeline that uses production keys, a branch created by the assistant can reach them. Keep production credentials out of jobs that run on unreviewed branches.
Revoking a connection does not undo what was already done. If you suspect the chat provider or your phone was compromised, revoke every connection, touch the stop file, check the remdesk journal for calls you do not recognise, and rotate any key that was inside the roots.
This article does not certify any product, including ours. Prices are planned, remdeskroot and tvijoterm are not deployed.
FAQ
Is it unsafe to use an AI chat on my phone for development at all?
No. It is safe enough when the chat works in a repository or workspace without secrets and cannot deploy. The risk comes from giving the phone connection rights over production.
Why not just tick "ask before every call" in the chat app?
That prompt lives in the chat app. It protects you from the assistant acting without you, but not from someone who holds the connector token. The server's limits and the boxes you granted are the real boundary.
Which remdesk plan should I take for phone work?
Free covers the phone: read, write and exec inside the workspace, 10 000 calls a month. Paid plans add deploy and full, which belong on a device you trust and a session you are watching.
Can I install remdesk or a terminal on my server by asking the chat?
Only if the chat already has the right to install software there. Do the first installation of remdesk yourself from a laptop or SSH session. tvijoterm is designed to arrive as an uploaded file installed by an action you approved with a ticket; even then the chat can only open or close its window, never pair a new device.
Facts checked on 4 November 2026 against the remdesk OAuth notes, the draft permission catalogue (prices planned), the FinEdg terminal specification and the tvijoterm design notes. AI-assisted editorial text prepared for Tvijo. tvijoterm and the root executor are not deployed.
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