Guides

Every guide follows the same shape: what you get, the concept in a few lines, the Agentfile, the commands, then a section on where enforcement stops, linked to the Known limitations. Read that section before you rely on a guide.

Start hereLink to this section#

• Agentfile in 5 minutes: write a first Agentfile, read what constle validate tells you, and avoid the fields that parse and do nothing.
• Sandbox an AI coding agent: run a CLI coding agent headless in a sandbox with one allowed API host and only the variables you declare.

Network and toolsLink to this section#

• Stop an AI agent exfiltrating data: an egress allowlist the agent cannot step around, and the channels it still leaves open.
• Lock down MCP servers: put every remote MCP server behind a gate, with a tool allowlist the agent cannot skip.
• Require human approval for MCP tool calls: hold a named tool call until a person approves it, and refuse it when nobody answers.
• Cap what an AI agent can spend: per-run and per-day caps on priced tool calls, and the spending they do not cover.

Containment and evidenceLink to this section#

• Contain prompt injection: assume an injected instruction succeeds, then limit what it can reach, call, spend and how long it runs.
• Give each AI agent a verifiable identity: one did:key per agent, a private key that never enters the sandbox, and what it signs.
• Verify what an AI agent did: read the signed audit log and check it offline, pinned to the identity you expect.

The concepts behind themLink to this section#

The guides link to these pages for the detail: Architecture, Network isolation, Human gates, Spend caps and metering, Audit log and verification, Identity and the Field reference.