Docs · console

PriestAltar console

The console is the same engine as the panel — the same conversations, the same tools, the same permission checks — used from a terminal. Open it from the PriestAltar desktop shortcut, or from any prompt with the priestai command.

Docs home Install The console Models & seats Permissions The fleet DevAd FAQ

Launching a model

Open from the website or your computer

The website’s Menu → Open PriestAI action offers a link to the installed PriestAltar console. On Windows, Setup registers the priestai://open/console app link for your user account. Existing installations need to run Setup again to register it. Your browser may ask before opening the app; visiting a page never launches it automatically.

If nothing opens, use the PriestAltar shortcut, or run priestai --console in a terminal where PriestAI is on your command path. Installation instructions explain how to set up the app.

Inside the console, /menu opens a numbered command palette; add a search term to narrow it, or press Enter to cancel. Use /theme night to change appearance, or launch with priestai --console --theme night. The six named themes are Night, Ember, Moss, Paper, Vigil and Reliquary; Dark and Light are also accepted. Set PRIESTAI_REDUCED_MOTION=1 to disable the animated pane loader.

Start a conversation

Choose a model to start a conversation, similar to ollama run works. The name goes through the one resolver — an installed model, a short name, an OpenRouter id, or the agent-mode word — and what answered is printed before the first prompt:

PriestAI> priestai launch qwen # installed model (a unique prefix is enough)
PriestAI> priestai launch claude # a paid cloud model, within your spending limit
PriestAI> priestai launch openclaw # agent mode: PriestAI's own tool loop
PriestAI> priestai launch anthropic/claude-sonnet-5 # an exact catalog id
PriestAI> priestai launch qwen --first "read the logs" # send a first message, then stay

When a name could mean more than one model — two qwen models installed, say — it refuses and names the candidates, so a model is never selected on your behalf.

Resuming with any model

A conversation is not tied to one model. Resume it and choose which model answers:

PriestAI> priestai resume # continue the newest conversation, current model
PriestAI> priestai resume claude # newest conversation, answered by Claude
PriestAI> priestai resume claude,qwen,openclaw # the first of these that answers, in order
PriestAI> priestai resume 20260903101422 qwen # a named chat (see /chats), answered by qwen

The comma list is tried in the order you typed it: the first model that answers is used. If none answers, the command refuses and lists every miss — it never silently keeps the old model. Inside a conversation, /model claude switches models the same way, mid-conversation, and the whole thread comes with you.

Task activity and messages

In a conversation, choose Activity to inspect recorded tool calls, worker progress and messages from other tasks. Changes opens captured file differences. Reviewing a difference does not apply it.

The shared command line is available on Windows and Linux:

priestai --cli task list --json
priestai --cli task new --title "Review the project" --workspace . --json
priestai --cli task inspect TASK_ID --json
priestai --cli task activity TASK_ID --json
priestai --cli task changes TASK_ID --json
priestai --cli task message TASK_ID TARGET_ID "Please review the changes" --request-id review-1 --json

Use task IDs returned by the list command. Reuse the same request ID and text to retry a message safely. Messages are delivered to another owned task in the same workspace; delivery does not start work or grant permission.

Command-line records belong to your operating-system account. Browser accounts have separate records. Goal budgets are requested limits unless enforcement is explicitly reported. Missing usage and file line counts remain unavailable.

Every command

CommandWhat it does
/model [name]Show the current model, or switch this chat to another — installed names, short names like claude, or OpenRouter ids, all through the one resolver.
/modelsEvery model on this machine and on any other machine of yours that answered the probe, with sizes.
/pull [model] [host]Download a model onto a machine — this one by default — and watch the progress. Ctrl+C stops watching, not the download; a bare /pull shows what is downloading right now. The target must be a machine the install knows (see /hosts).
/hostsEvery machine this install can put a model on, probed live: which server answered, whether it can download models, and how models get there if it cannot — plus the cloud tier with its key status and this month's spending.
/catalog [word]The installable catalogue — the Ollama library and Hugging Face GGUF files, with sizes and an uncensored marker; a word narrows it. This is what /pull can install.
/adopt <host>Add a second machine running Ollama as a model server. It is checked at the moment you add it, so an address that does not answer is refused rather than saved.
/key [value|clear]The OpenRouter key. Bare, it shows the status — stored or not, where it came from, the last four characters — never the value. A value stores it; clear removes it. Stored through the same code path the panel uses.
/seatsWhich model is assigned to each role, and what each one measured.
/bareToggle bare mode (on by default): nothing is added to what you typed.
/toolsThe tools available in this conversation.
/skills [id]The skills on this machine; name one to switch it on or off for this conversation (at most 4 at once).
/project [name]Which project's shared notes read_notes and post_note use. No name clears it.
/newStart a fresh conversation.
/chatsRecent conversations.
/open <id>Resume one by id.
/attach <path>Attach a file. A local path, or a UNC path on another machine you own — a remote file is read in place, without maintaining a separate attachment copy.
/attachedWhat this conversation has attached.
/peersEvery other machine of yours, and whether it is online.
/browse <host> [p]List a directory on one of them over SMB.
/perm [mode]What the assistant may do without asking — see Permissions.
/clearClear this conversation's messages (the record of tool runs is kept).
/costThis month's paid-model spending against your limit — the same two numbers as the Money panel. Local inference has no provider fee.
/toolviewOpen the panel with the agent's picture-in-picture window, which shows what the agent is looking at while it works.
/updateAsk the panel for the newest release and, with your yes, install it through the same verified path the app uses; if the new build fails to start, the previous build is restored automatically.
/quitExit the console.

Ctrl+C cancels the current turn

Ctrl+C cancels the current turn and returns you to the prompt; the conversation is kept. Press it again at an empty prompt to exit.

Shell commands

A line starting with ! runs a shell command where you are — ! git status, say — with the output streaming as it runs, so you do not have to leave the console to run a build.

A known limit: bare mode is the default, so the model receives only what you typed. Start with priestai --briefed (or the console flag of the same name) if you want the app's briefing text sent along with your messages.