Guide
Keep API keys out of your agent
An AI agent can read every key on the machine it runs on. Once the keys are in Actori, the agent calls your services without holding one. The first service takes about ten minutes.
Where the keys are now
A coding agent with a shell reads files, which means that whatever a key in those files can do, the agent can do as well, with no rule in between and no record of it afterwards. Which files? Usually these.
- ✓
.mcp.jsonand~/.claude.json: the tokens in each MCP server'senvorheaders. - ✓
.envfiles in the project. - ✓Command-line logins, such as
~/.aws/credentials. - ✓A shell profile that exports a token.
What changes with Actori
The agent gets one address, https://agents.actori.ai/mcp, which it signs in to as you. Actori holds the credential of each service and makes the call itself, so what the agent gets back is the result of the call and never the credential that made it.
- ✓A key never passes through the chat. You sign in to a service, or paste its key, on Actori's page in your browser. Actori refuses a key an agent tries to hand it.
- ✓If a service echoes its token back in an answer, Actori removes it before the agent reads the result.
- ✓AWS can connect by role. You add a trust policy to a role of yours, which leaves no access key to store.
- ✓Every call is checked against your rules first. And tools that delete or send wait for your approval from the moment you add them.
Move one service
Start with the key that could do the most damage. The steps below use GitHub; any other service goes the same way.
- ✓Connect Actori. Paste
Set up Actori for me: https://actori.ai/into your AI and sign in when it asks. - ✓Ask it:
Connect GitHub through Actori.It adds the connector and gives you a link. You sign in to GitHub there, in your browser. - ✓Ask it for the tools you need:
Give yourself the tools to read issues and open pull requests.You approve the request in Actori. The tools appear in a new conversation. - ✓Remove the old GitHub server from
.mcp.json. Then delete its token from the machine. - ✓Revoke the old token at GitHub. Why revoke a token that never leaked? The agent could read it, which is reason enough to treat it as used. And do not make a new one for the agent.
Check it worked
- ✓Ask the agent to print your GitHub token. It has none to print.
- ✓Search the project and your home folder for the first characters of the old token. Nothing should match.
- ✓Ask the agent to do one thing that writes, such as opening an issue. The call shows in Actori with the rule that decided it.
What this does not cover
- ✓A key that stays on the machine stays readable. Actori protects the services you connect through it and nothing else.
- ✓Commands the agent runs in your shell, such as
terraform, a cloud CLI or a database client, never reach Actori: they run on whatever logins the machine has. So give those a read-only login and send the writes through Actori. - ✓Inside your rules the agent still acts as you. Say what must never happen and what must wait for you. Your AI writes the rule from that, and you approve it.
Links
- –Set up Actori — Connect your AI app, step by step.
- –Security — How Actori stores credentials and who can reach them.
- –The agent deleted production because the token let it — What one such key did, and what to change.
- –Home — What Actori does.