Bloom is a design canvas where people and agents work together. These plugins wake your agent's conversation when someone in Bloom tags it, or answers a question it asked, even while the conversation is idle. Install once per computer; after that, every conversation that uses Bloom listens, with nothing to start in each one.
| Agent | Install | Update |
|---|---|---|
| Claude Code | claude plugin install bloom --marketplace BloomLabsInc/agent-plugins |
claude plugin marketplace update bloom && claude plugin update bloom@bloom |
| Codex | codex plugin marketplace add BloomLabsInc/agent-plugins && codex plugin add bloom-wake@bloom |
codex plugin marketplace upgrade bloom && codex plugin add bloom-wake@bloom |
npx @bloom/cli mcp setup claude and npx @bloom/cli mcp setup codex connect Bloom's MCP server
and install the plugin in one step. Then:
- Claude Code loads a newly installed plugin when a session starts: type
/reload-pluginsin a running one, or start a new one. - Codex asks you to trust the plugin's hooks the first time it starts after an install or update. Choose "Trust all and continue".
The plugins run on macOS and Linux (WSL included) with Node 22 or newer.
Each plugin holds the agent's hooks, a short POSIX shell launcher (bin/bloom-hook), Claude Code's
waiter (bin/bloom-wait), and cli-version, the @bloom/cli release the hooks run. The launcher
runs that release with npx -y --prefix ~/.bloom @bloom/cli@<version>, so it never picks up a
project's own copy.
| Hook | When | What it does |
|---|---|---|
PostToolUse, MCP tools only |
After the agent calls an MCP tool | When a Bloom answer carries a one-time code for this conversation, swaps it for a token that can only read this conversation's Bloom tags, and hands the conversation to the machine's one watcher. Results without a code end in the shell script, with no network request. |
SessionStart and Stop (Claude Code, in the background) |
When a session starts, and when Claude finishes a turn | Waits for the conversation's tags and answers. When one arrives it exits with code 2, which wakes the session, and Claude reads it from Bloom. A conversation Bloom never connected stops waiting after two minutes. Nothing runs in claude -p. |
SessionStart and UserPromptSubmit (Codex) |
When a thread opens, and when you write in it | Starts the thread's watcher again if it stopped. Codex Desktop gets each tag steered into the thread; Codex CLI gets it through codex queue. |
A code is kept only when the session it names is this conversation's own, and only for Bloom's own servers. Any other MCP server's result is ignored.
- Once, on first use:
npxdownloads@bloom/clifrom npm into~/.bloom. - For connected conversations only: the CLI talks to the Bloom deployment that issued the code. It exchanges the code for its token, renews the token every hour, reads the conversation's tags, and sends a heartbeat so Bloom shows the conversation as connected. It never sends your prompts, the agent's replies, tool results or files.
Everything lives in ~/.bloom: the CLI, the watcher's socket, each connected conversation's link
and token (owner-only files), and two logs, watch.log and plugin.log. Neither log ever contains
a token or a code.
- Pause: set
BLOOM_PLUGIN_DISABLED=1in the agent's environment, or disable the plugin (claude plugin disable bloom@bloom). - Uninstall:
claude plugin uninstall bloom@bloom, orcodex plugin remove bloom-wake@bloom. Then delete~/.bloomif nothing else on the computer uses Bloom's CLI.
This repository is published from Bloom's main repository at each @bloom/cli release, so a pull
request here is applied there. Report a security problem privately through
Bloom support, not in an issue.
Apache License 2.0; see LICENSE.