● attached · 4 hosts · 6 sessions
You Are The Ultimate Agent.
The models are good at typing. You are the one who knows what should be
built. The Architect gives you one console for every harness you already run —
attached over SSH, inside the screen and tmux sessions you
already keep alive.
No orchestration layer. No opinion about your workflow. Just direct hands on every agent you have running, from whichever device is in front of you.
One line per job: project, last output, elapsed time. When a session rings the terminal bell, the badge lights and decays over sixty seconds — so you can tell an agent that wants you now from one that wanted you an hour ago, without opening anything.
Stop texting in circles.
You have had this conversation. You describe the change. It rewrites the wrong file. You correct it. It apologizes and rewrites the wrong file again. Forty minutes gone to a chat window that never had the context to succeed, running somewhere you cannot reach.
The fix was never a better prompt box. It is you, at the controls, with the real terminal in front of you — picking the harness, picking the model, reading the actual output, and stopping the run the moment it goes sideways.
- ssh
- Keys are already on the boxes that matter to you.
- tmux / screen
- You keep sessions alive across disconnects, on purpose.
- a harness
- Claude Code, Codex, Aider, OpenCode, or the one you wrote yourself.
- judgment
- You decide which model gets which job. Nothing here decides for you.
If that list reads as a foreign language, The Architect is not for you.
This is not an onboarding funnel with a terminal theme. It assumes you already run multiple harnesses and multiple models across real machines, and that your only missing piece is a decent way to reach them all from wherever you happen to be.
There is no tutorial, no starter template, and no assistant to hold your hand. It attaches to what you have already built and gets out of the way.
Leave the amateurs in the dust.
Swipe to the next session.
A real ANSI terminal — truecolor, correct UTF-8, live PTY resize, mouse
reporting, Kitty keyboard protocol. On a touchscreen, swipe left or right to move
between your screen or tmux windows. On desktop you get
the same move on buttons, so drag-to-select keeps working.
> rewrite the bell decay to use the app clock ● Read src/core/notifications (214 lines) ● Edit src/core/notifications +18 -6 ● Bash make check No issues found. ● Done in 41s · 2 files changed ▮
aider> /add collector/ingest.py Added collector/ingest.py to the chat. tokens: 12.4k sent, 891 received ● Apply edit to collector/ingest.py? (y/n/skip) ▮
$ pytest -q tests/ ................................ F..... FAILED tests/test_sync.py::test_upsert_by_device AssertionError: expected 2 entries, got 3 38 passed, 1 failed in 4.12s ▮
$ tail -f /var/log/deploy.log 14:02:11 pulling origin/master 14:02:14 build ok (48.2s) 14:03:02 restarting workers … 14:03:04 4 workers up ▮
Drag it. Or use the arrows — same as the app.
Exact-name attach
GNU screen matches session names by prefix, which is how people end up typing into the wrong agent. The Architect matches the name exactly and attaches by pid-qualified id, or creates fresh. It never guesses.
The session outlives the app
Sessions live on the remote host, not in the client. Close the app, lose signal, swap devices — the run keeps going and you reattach to it exactly where it is.
Same console, every screen.
Not a stripped-down companion app on the phone and a real one on the desktop. The same terminal engine everywhere — the navigation adapts, the capability does not.
Set the environment before you connect.
Every serious harness is configured through its environment. Which endpoint it talks to, which model it asks for, which key it authenticates with, which proxy it goes through — all of it arrives as variables, and all of it has to be right before the process starts, because none of it can be changed once the tool is running.
So you set it on the agent, not in a .bashrc you have to remember
to edit on four different machines. The Architect exports your variables into the session,
then runs your start script, then launches the tool — in that order, every time.
One place, per agent, under your control.
# exported before the tool starts ANTHROPIC_BASE_URL=https://llm-gw.internal:8443 ANTHROPIC_API_KEY=sk-ant-•••••••••••• ANTHROPIC_MODEL=claude-opus-5 OPENAI_BASE_URL=http://workstation.lan:11434/v1 HTTP_PROXY=http://proxy.lan:3128 NO_PROXY=localhost,127.0.0.1,.internal TERM=xterm-256color COLORTERM=truecolor
Point a harness anywhere
A base URL is just a variable. Send one agent through your company gateway, another at a model running on the box down the hall, a third at a staging endpoint you are evaluating — same tool, three agents, no reconfiguration between them.
Keys per agent, not per machine
Each agent carries its own set. A host shared across projects does not need one
global key in a shell profile, and rotating a key is one edit in one place instead of
an ssh round trip to every machine.
A start script for the rest
Whatever variables cannot express — activate a virtualenv, change into the project directory, warm a cache. It runs after the exports and before the tool, so the harness comes up in the state you meant. An end script runs when you disconnect.
One place. Full control.
Your config syncs through your own server.
There is no sync service, because there is no server of ours to sync to.
The Architect writes your agent and project configuration — environment blocks
included — as JSON over SFTP into ~/.thearchitect/ on a host you
already own, and reads it back on the next device. That is the whole mechanism.
Push
Export agents and projects to any host you can reach. Hosts, ports, session managers, tools, env vars, start and end scripts.
Pull
Point a new install at the same host. The first-run wizard lists what is there and imports what you pick.
Never SSH credentials
Your SSH passwords and private keys are excluded from the transfer by design. They stay in the device keystore and are never written to a host — not even the host you are exporting to. Your environment variables do travel, into your own home directory on your own machine, because that is what makes a new device work on arrival.
Same idea for keys: the app generates an RSA-2048 and an ECDSA P-256
identity pair on first launch, copies the public line to your clipboard, or appends it to a
host's authorized_keys over a password session — so you can move to key
auth without leaving the app.
It does not care who you are.
The Architect has no idea what your name is, and has no way to find out. There is nothing to log into and nothing that phones home.
- □ No account, no sign-in, no profile The app ships without a login screen. There is no identity to create, verify, recover, or breach.
- □ SSH credentials never leave the keystore Private keys and passwords go to the platform keystore — Keychain, Credential Manager, libsecret — and are read only to open your own connection. Never the app database, never a config file, never written to a host, never sent anywhere.
- □ Traffic goes to your hosts and nowhere else Every connection is a direct SSH session from your device to your machine. No relay, no proxy, no broker in the middle.
- □ Licensing is offline Pro is unlocked by a platform-locked key validated on-device. No activation call, no seat server, no license check that can fail while you are on a plane.
- □ No telemetry Nothing measures how you use it. If you want a record of a session, turn on traffic logging and it writes to a file on your own disk.
Take the free one.
One agent and one project, free forever, with every feature on this page working. It is enough to attach to a real host and decide for yourself. Pro lifts the limits when you are ready.
- 1 agent, 1 project
- Full ANSI terminal, swipe between sessions
- Bell badges and live work status
- SFTP config sync
- Device identity keys
- Unlimited agents and projects
- Offline platform-locked key
- License travels with your SFTP config
- Everything in Free
Direct in the US; through the app stores everywhere else.
- Every Pro capability, no charge
- Key issued directly to your inbox
- AppImage build
- No storefront to go through
Windows, macOS, and Linux today. iOS and Android coming soon. Linux keys are issued by us directly, anywhere in the world.