Using VyreDraft

The vyre command

vyre is Vyre in a terminal, on your Mac or on your own server over SSH. Every command is a call to Vyre's background service on that machine, so the terminal, the Vyre app, your phone and Lumen never disagree about what is true. This page is organised by task. Every command and its usage line is in the CLI reference; vyre help prints the same list, and vyre help <command> shows one command's usage and flags.

Where to run it#

vyre update               # fetch the newest stable release, check it, back up, rebuild and recreate the box
vyre update --rollback    # put the previous release back
vyre logs                 # follow Vyre's output
vyre uninstall            # stop Vyre and take it off this server; it asks whether to delete the data

vyre box            # which box, and whether it answers
vyre box update
vyre box backup

Start Vyre and check on it#

vyre up        # start Vyre; print your box's address
vyre status    # is it running, which version, how many modules
vyre modules   # every module and whether it started
vyre down      # stop Vyre

  vyred running · 0.2.0 · box · pid 412 · up 3600s

vyre up is safe to run any time. After an upgrade it restarts an older Vyre. When it has nothing left to set up, it prints the same "Vyre is ready" block every time.

If this happens

vyre status says a module failed#

Run vyre modules. A failed module shows its error on its line; Vyre keeps running without it.

Open your projects#

vyre with no arguments opens the home: every project, a new session without a project, and your agents. Arrow keys move, typing filters, Enter opens. Picking a session resumes it in Claude Code. Inside a project's folder, that project is preselected. Piped, it prints the list and exits.

vyre projects             # every project
vyre open juniper-intake   # what a project's threads are told, and its threads
vyre new                  # make a project by picking sessions

Find something you said#

vyre recall "retainer letter for Juniper Studio"
vyre threads juniper       # sessions on this machine that mention juniper

vyre recall searches every past session by meaning and by words. Its flags:

vyre recall with no words says how many sessions are indexed. vyre index indexes new sessions now instead of waiting.

Resume or start a thread in Claude Code#

vyre resume <thread>      # open it where it ran, with its project's brief
vyre start                # a new thread in this folder's project
vyre context              # what a new thread here would be told

These hand your terminal to Claude Code until it exits. They load Vyre's hooks and tools with --plugin-dir. To get the same in a claude you start yourself, install the Vyre plugin: see Vyre in Claude Code.

Drive a running session#

Headless threads are sessions Vyre runs in the background on your server, so they outlive every terminal, browser tab and closed laptop.

1. Start one:

   vyre threads start --project juniper-intake "draft the intake checklist"
   

   You should see

     started 3f9c2a71-...  juniper-intake · ~/Vyre/projects/juniper-intake
     vyre threads watch 3f9c2a71
   

   Without --project, it starts in the current folder (--cwd picks another). --name, --model and --provider are optional. --provider codex (or grok, openrouter) starts it on that provider's account instead of Claude's.

2. Watch it. Its recent history prints, then what it does next, until it stops or you press Control-C:

   vyre threads watch 3f9c
   

3. Send it more words, or answer what it asks:

   vyre threads send 3f9c "also add a conflicts check"
   vyre threads asks                    # what is waiting on you
   vyre threads answer <ask> allow      # or deny, with an optional message
   

4. Stop it: vyre threads stop 3f9c.

To change a running thread's model, vyre threads model 3f9c sonnet (or a model id). To move the thread to another provider between turns, which is what the model picker in Chat does:

vyre call threads.switch '{"thread":"3f9c...","provider":"codex"}'

The thread says "Switched to Codex" and carries on with the same folder and files. It is refused while a turn is running. A provider with more than one account also needs "account". Sessions says what carries over.

vyre threads list shows headless threads from the last day (--all for every one). A thread id can be shortened to its first four or more characters.

Only one surface types into a session at a time. If another holds it, send says who; vyre threads lease <thread> takes it and vyre threads release <thread> gives it back.

If this happens

"This session is open somewhere else"#

send resumes a stopped session, or one you ran in a terminal, only when no other process has it open. Close it in the other terminal, or type there.

Ask an agent#

vyre agents                          # every agent and what it is doing
vyre agents ask juno "what is left on the intake form?"
vyre agents threads juno
vyre agents usage juno               # turns, time, tokens, spend, rate limits

vyre agents ask waits for the answer (up to ten minutes). If the agent stops on a permission question, it prints the ask and the vyre threads answer line to answer it.

vyre agents create kit --projects juniper-intake makes an agent; see Agents for every flag, and for vyre agents computer kit to look after an agent's own computer.

Ask what memory knows#

vyre memory                    # what memory holds
vyre memory "Northwind Bakery" # everything about one thing
vyre why "<fact>"              # the turns a fact came from

To correct memory, see Memory.

Use a credential without seeing it#

vyre vault put stripe-key                  # prompts without echo
vyre vault run stripe-key -- npm run deploy

vyre vault run puts the value in one command's environment, named after the item (STRIPE_KEY), and scrubs it from the command's output. VAR=item.field picks the variable name and field. Everything else the vault does is in Vault.

Script anything#

Every tool Claude and the surfaces can call is callable from the terminal:

vyre tools                                   # every tool, with a line on each
vyre call projects.list
vyre call system.echo '{"text":"hi"}'

vyre call prints the tool's data as JSON, or an error with its code. A tool that needs a person (an approval, a vault reveal) asks you to prove you are here: Touch ID when Vyre offers it, or a code typed in this terminal. --tty asks for the code in the terminal.

For machine-readable output, add --json: most commands take it, and vyre help <command> shows whether one does.

What it will not do#

• It will not take a secret on the command line, where shell history and Claude's transcript would see it. vyre vault put prompts or reads stdin.
• It will not approve anything for a script or for Claude's own shell. Approvals need a person at a terminal: a process with no terminal is refused.

How Vyre runs its sessions#

vyre sessions shows and sets how the sessions Vyre starts run: vyre sessions status (driver, sign-in, Claude Code, Agent SDK), vyre sessions models (the model per kind of session) and vyre sessions prompt (the system prompt at three levels, with history and undo). Changing a model or a prompt is refused from inside a Claude session: do it from the Vyre app or a plain terminal. Sessions has the rest.

Next#

• CLI reference, every command and usage line.
• Projects and threads, what the project commands work on.
• Vyre in Claude Code, the plugin and vyre statusline.
• Tools reference, everything vyre call can run.