A Vellum Assistant plugin that lets any agent manage a Vellum Assistant, including another Vellum Assistant.
Built for Claude Code, Codex, OpenClaw, Hermes, and any other agent that can call tools or run a CLI. Install it on your Vellum Assistant and it gains structured tools for controlling another instance. Or use the vellum CLI directly from any agent with bash access.
Agents are increasingly managing other agents. A CI pipeline wants to restart an assistant after a config change. A coding agent wants to install a plugin on the user's assistant. A "supervisor" assistant wants to check on and coordinate other assistants. Today every one of these workflows means hand-rolling vellum CLI calls and hoping the agent figures out the right incantation.
This plugin packages that knowledge into typed tools and a skill that teaches the agent the management workflow, guardrails, and common patterns.
The plugin exposes tools that wrap the vellum CLI and IPC layer. An agent calling these tools can:
- Check status. Is the daemon running? What version? What workspace? What processes?
- Send messages. Start a conversation or ask a question and get a response back, via IPC.
- Manage plugins. List, install, uninstall, and upgrade plugins on the target assistant.
- Manage schedules. List, create, and delete scheduled tasks.
- Manage routes. List and configure inbound routes.
- Manage credentials. List credential names (never values) and check connectivity.
- Read logs. Pull recent daemon logs for debugging.
- Restart. Stop and start the daemon.
| Surface | What | Why |
|---|---|---|
| Tools | CLI-wrapping tool definitions the model can call | Core value. The agent calls these to manage the target assistant. |
| Skills | SKILL.md teaching the management workflow |
Tells the agent when to use each tool, what order to do things, and what to watch out for. |
| Hooks | init only |
Validates the target assistant is reachable at boot. |
No user-prompt-submit, pre-model-call, or other transform hooks. This plugin does not intercept or modify the host assistant's own conversation flow. It only adds outbound management capability.
┌─────────────────────────┐ ┌─────────────────────────┐
│ Host Assistant │ │ Target Assistant │
│ (this plugin installed) │ │ (being managed) │
│ │ │ │
│ tools/vellum-*.ts │ │ vellum daemon │
│ ↕ model calls │ │ ↕ IPC socket │
│ src/vellum-cli.ts │──CLI──▶ │ vellum-cli.sock │
│ ↕ subprocess │ │ vellum-ipc.sock │
│ vellum CLI │ │ │
└─────────────────────────┘ └─────────────────────────┘
The host assistant (where this plugin is installed) calls the vellum CLI as a subprocess to manage the target assistant. For local management (same machine, same workspace), the CLI talks directly to the daemon's IPC sockets. For remote management, the CLI can target a remote workspace over SSH.
The target can be the host itself. An assistant can manage its own daemon, plugins, and schedules through this plugin. The most common use case is managing a different assistant instance.
# From the marketplace (once published)
vellum plugins install vellum-manager
# From GitHub (development)
vellum plugins install https://github.com/vellum-ai/vellum-pluginRestart your assistant after install.
| Field | Required | Default | Description |
|---|---|---|---|
targetWorkspace |
No | Current workspace | Path to the target assistant's workspace. Defaults to managing the host itself. |
sshHost |
No | — | SSH host string (user@host) for remote management. When set, CLI calls are wrapped in ssh. |
If neither is set, the plugin manages the assistant it's installed on.
The vellum CLI is the underlying interface. Any agent with bash access (Claude Code, Codex, OpenClaw, Hermes) can call it directly:
# Check if the daemon is running
vellum status
# List installed plugins
vellum plugins list
# Send a message and get a response
vellum chat "What's on my schedule today?"
# Install a plugin
vellum plugins install https://github.com/owner/my-pluginThe SKILL.md in this repo doubles as a reference for non-Vellum agents. It documents the full management workflow, tool selection logic, and guardrails. Point your agent at it.
- Credential values are never exposed. The
vellum_credentials_listtool returns names and connectivity status only.vellum credentials revealis deliberately not wrapped. - Restart and plugin install are high-risk actions. These tools declare
defaultRiskLevel: "high"and require user confirmation on first call. - Remote management requires explicit SSH config. The plugin will not SSH anywhere unless
sshHostis set in config. - The agent is not the guardian. Per Vellum's guardian model, the plugin does not expose tools for mutating the contact/trust graph. Those writes are gateway-only and never proxied through an agent.
Pre-alpha. The repo is scaffolded but not yet functional. The README exists to align on scope before implementation.
MIT