Agents
An Agent is a member of your Organization that is not a person. Create one when you want an AI tool such as Claude Code, Cursor or Codex to read your Services or deploy them through the MCP endpoint.
Applies to: Owner or AdminEvery plan
An Agent has its own role, its own grant to the Projects, Environments and Services it may reach,
and its own agent keys. It acts only through the MCP endpoint, the console’s address plus
/mcp, and every call it makes, allowed or refused, is a row in the audit log.
What you can do
- Create Agents with the Reader, Operator or Caller role.
- Grant each one only the Projects, Environments and Services it needs.
- Issue, rotate and revoke its agent keys.
- Suspend an Agent at once, and resume it later.
- Copy ready-made configuration for Claude Code, VS Code, Cursor and Codex CLI.
Agent roles
| Role | What it can do |
|---|---|
| Reader (default) | Reads what it is granted: Services, Environments, Deployments, Deployment logs and Domains. Changes nothing. |
| Operator | Everything a Reader can, and deploys, starts, stops and cancels Deployments of what it is granted. Cannot change settings, variables or access. |
| Caller | Calls the AI models it is granted through the Gateway, and nothing else. Gets no tools over MCP. |
Create an Agent
Open Agents
Open Settings > Agents and press Create an Agent.
Each Agent shows its role, what it can reach, its keys and when a client last connected. Name it and pick a role
Give it a Name that says what it is for, such as
deploy-botorstaging reader; it is how the audit log names it. Pick a Role: Reader is selected by default.Grant access
Under Access, tick what it may reach, for example the
storefrontProject’sstagingEnvironment. Access starts empty, and granting a Service grants its Project and Environment too. You can save it with nothing granted and grant it later.Choose when its first key expires
Under First key expires, choose In 30 days, In 90 days (the default) or In a year. Press Create Agent.
Copy the key
The first key is shown once. Copy it into your MCP client’s secret storage now.
Connect an MCP client
On the Agent’s row, press Connect. The panel holds the configuration for Claude Code
(.mcp.json), VS Code (.vscode/mcp.json), Cursor (~/.cursor/mcp.json), Codex CLI
(~/.codex/config.toml) and any other client. Each sends the key as
Authorization: Bearer <agent key> to your console’s address plus /mcp.
Step-by-step client setup is in Connect an MCP client.
Rotate, suspend or delete an Agent
- Add key issues another key with its own expiry. To rotate, add a key, switch the client to it, then Revoke the old one. An expired key shows Remove.
- Suspend refuses its next call through the MCP endpoint, and every one after, until you press Resume.
- Change role and Manage access change what it can do and reach.
- Delete removes the Agent, its access and all its keys. Its audit trail stays.
What you cannot do
Related pages
- Connect an MCP client: set up Claude Code, Cursor, VS Code or Codex.
- MCP tools: every tool an Agent can call.
- Guard rails: masking, rate limits and refusals.
- Audit log: every call an Agent made.
- Members and roles: how grants work for people.
- Cloudgate for AI agents.
Questions
Can an Agent apply a Plan or change settings?
No. An Agent can ask for a Plan but never applies one, and never changes settings, variables, access, Uptime checks or thresholds.
Does an Agent see my secrets?
No. Every tool result, logs included, passes through a mask that replaces stored secret values with dots.
What happens when the person who made an Agent leaves?
Nothing. An Agent outlives whoever created it; it belongs to the Organization.
Can I use a personal key for an agent instead?
No. The MCP endpoint refuses personal keys. Create an Agent and use its agent key.
Something missing or wrong on this page? Email support@octabyte.io, or start a free trial and try it in the console.