Commands

Celly registers guild-scoped slash commands (instant updates) plus a ! shell prefix. /project uses subcommands.

Project management

Command	Access	Behavior
`/project add name:<name> path:<path>`	owner	Register an existing directory under `PROJECTS_ROOT`; runs the create saga.
`/project create name:<name>`	owner	Create a directory under `PROJECTS_ROOT`, then add it.
`/project list`	authorized	List projects with sandbox status and health.
`/project status name:<name>`	authorized	Status, port, health, and session count (ephemeral; password masked).
`/project start name:<name>`	owner	Wake the sandbox and supervised server (recreates it if missing).
`/project stop name:<name>`	owner	Stop the supervised server, then the sandbox.
`/project remove name:<name> confirm:<name>`	owner	Typed confirmation; remove sandbox, rows, and channel.

Owner-only means the guild owner, or a member with OWNER_ROLE_ID (or Administrator/Manage Server). list and status are available to any authorized member. Access is decided by ACCESS_ROLE_ID / BLOCK_ROLE_ID, with block checked first.

Sessions

Command	Access	Behavior
`/new [prompt]`	authorized	Start a new thread/session in this project channel.
`/resume`	authorized	Pick a past session (ephemeral select) and resume it in a new thread.
`/abort`	authorized	Abort the current run in a thread, or every active run in the project channel.
`/model`	authorized	Choose the model for this thread (two-step provider → model).
`/agent`	authorized	Choose the agent for this thread.

Messages and shell

Input	Where	Behavior
Plain message	project channel	Creates a thread, a session, and sends the prompt.
Plain message	thread	Continues that session (queued if a run is active).
`!<command>`	channel or thread	Runs `bash -lc <command>` inside the project’s sandbox and posts the output.

Text-like attachments on a message are size-capped, written to a validated .celly/inbox, and referenced in the prompt. Shell output is truncated and chunked across up to three messages with a total-length footer.

!<command> runs with the same privilege as the agent inside the microVM. It cannot run host commands, but it can do anything the sandbox allows. Treat it as running code.

Deferred to v1.1

Commands: /project restart, /share, /diff, /undo, /redo, /context-usage. Features: worktree-per-thread, /btw forks, queue UI (. queue), permission-approval buttons, question rendered as Discord components, voice messages, image attachments, OpenCode web UI, tunnels/screenshare, multi-guild, cloud sandboxes, --clone sandbox mode, OAuth subscription login, and Linux/macOS deployment docs. See the architecture reference for what happens under the hood when you run these.

Get started

Guides

Reference

Project

Project management

Sessions

Messages and shell

Deferred to v1.1

​Project management

​Sessions

​Messages and shell

​Deferred to v1.1

Project management

Sessions

Messages and shell

Deferred to v1.1