Everything you need to keep agents in check

Gatesi runs on Cloudflare's global network with a Telegram bot. There are daily limits (see the FAQ).

• Phone approvals. Risky requests reach you in Telegram, and in Slack if you like, with a plain summary, a risk level and the exact details. Approve, deny, or allow for a set time. If nobody answers, the request is refused after the wait you choose. An approval works once, for exactly the request you saw.
• Rules. Allow, ask or block tools and actions. Set cost limits, pause a runaway agent, and pause everything if nobody checks in for a set time.
• Vault. Store credentials encrypted. Gatesi makes each HTTPS API call for the agent and adds the key, so the agent never sees it.
• MCP proxy. Put any remote MCP server behind Gatesi. New or changed tools stay hidden until you approve them, and every call follows your rules.
• Two-member approval. Your biggest actions need a second approver. Only one-time approvals count.
• Honeytoken traps. Plant fake tools or fake secrets. If an agent touches one, you are alerted and the agent can be stopped on the spot.
• Tamper-evident record. Every agent step goes into a hash chain. Replay any run, and test new rules against old runs before you switch them on.
• Signed receipts. Sign a run with Ed25519. Anyone can check the receipt offline with our open checkers.
• Agent records, members and connections. Publish a read-only page about an agent's last 30 days. Invite up to 20 members, each with a separate account. Connect agents with our Python and JavaScript kits, a Claude Code hook, an MCP endpoint, framework wrappers or plain HTTP.

Accounts for other people

The admin can invite people, each with a separate account: their own agents, keys, vault, rules, runs and Telegram. Up to 20 people. The admin can see everything and can switch anyone off.

For big actions, you can ask a second member to approve as well.