Skip to main content
A managed agent is an agent with a core — a blessed runtime that OpenComputer knows how to configure, connect, and extend. Instead of managing snapshots and entrypoints yourself, you pick a core and build up from there. 
ManagedAgent = Core + Channels + Packages + Secrets

​
Cores

A core is the AI runtime that powers your agent. When you create a managed agent with --core, OpenComputer boots a sandbox with that runtime pre-installed, an LLM key configured, and a gateway ready to accept channels. 
oc agent create my-agent --core hermes
# or
oc agent create my-agent --core openclaw
CoreRuntimeWhat you get
hermesHermes AgentSelf-improving AI agent with skill creation, memory, multi-channel gateway, browser, code execution
openclawOpenClawMulti-channel AI gateway with rich plugin SDK, native MCP support, hot-reload config, 6 built-in channels + 20 plugin channels
Without --core, the agent uses the raw sandbox path — you provide your own snapshot and entrypoint via the API. Managed and raw agents coexist on the same platform.

​
Channels

A channel is a messaging platform connected to your agent. You connect and disconnect channels. 
oc agent connect my-agent telegram --bot-token <token>
oc agent disconnect my-agent telegram
ChannelHow it works
telegramRegisters a webhook with Telegram. Messages flow through the OpenComputer gateway to your agent’s sandbox. The agent replies directly to Telegram.
Channels are stateless — disconnecting removes the webhook and config. Nothing to clean up. When you connect a channel, the platform:
  1. Stores the channel secret (bot token) in the agent’s secret store
  2. Registers the webhook with the channel provider
  3. Configures the core for webhook mode
  4. Restarts the gateway

​
Packages

A package extends what the agent can do beyond its core. You install and uninstall packages. 
oc agent install my-agent gbrain
oc agent uninstall my-agent gbrain
PackageWhat it adds
gbrainPersonal knowledge system — 34 MCP tools for pages, vector search, graph links, timeline, files. Backed by a managed Postgres database.
Packages can be stateful. When you install gbrain, OpenComputer creates a managed Postgres database for that agent. The database lives outside the sandbox and survives instance restarts, uninstall/reinstall cycles, and even agent recreation. Uninstalling a package removes the wiring (MCP config) but preserves the data by default.

​
Putting it together

Three commands to a personal AI agent on Telegram with persistent memory: 
# Boot an agent (pick a core)
oc agent create my-agent --core hermes   # or --core openclaw

# Connect Telegram (create a bot via @BotFather first)
oc agent connect my-agent telegram --bot-token <token>

# Install persistent memory
oc agent install my-agent gbrain
The user journey is the same regardless of core. The adapter handles the differences — Hermes uses YAML config + gateway restart, OpenClaw uses JSON5 config + hot-reload. See the Create a Hermes Agent guide for a step-by-step walkthrough, and the gbrain package page for details on persistent memory.

​
The manual path

Managed agents are sugar over OpenComputer sandboxes. You can always drop to the sandbox level: 
# Shell into any agent's sandbox
oc shell my-agent

# From there: edit config, install software, debug, run native commands
# Hermes:
cat ~/.hermes/config.yaml
hermes skill list

# OpenClaw:
cat ~/.openclaw/openclaw.json
openclaw mcp list
The managed path and the manual path coexist. Channels and packages that you wire manually (via oc shell) won’t appear in the managed state, but they work fine.