Integrations · every surface, governed
One authority plane. Many gates.
Intended decides and proves; your agent acts. Every integration below sits inside your own agent or environment, asks Intended for a decision before an action, and carries the signed Authority Token back. Intended decides and signs. It never holds your keys or calls your systems.
02 · Decide-only integrations
guard() · authorize()
Wrap the action in your agent’s code. The SDK submits the intent, waits on escalations, and returns the decision with its token. Your code performs the action on ALLOW.
SDK docs →Customer-run MCP proxy
Runs in your environment between your MCP client and an upstream MCP server. It asks Intended for a decision before forwarding each tools/call; denied calls never leave your gateway.
MCP gateway →mcp.intended.so
Authority operations as MCP tools — submit an intent, verify a token, fetch a receipt, list policies, explain a decision. Your assistant still performs the action itself.
MCP docs →OpenAI live · Anthropic · Bedrock · Vertex · NIM (beta)
One env var on the box where your agents run. Each model tool call is evaluated before your agent can act on it. OpenAI is live-proven; the other providers are in beta.
Gateway →Pre-tool-use governance
A hook that asks Intended before Claude Code runs a tool. ALLOW lets the tool run on your machine; DENY and ESCALATE stop it there.
LangChain · Vercel AI SDK · OpenClaw
SDK integrations into the agent framework you already use, plus the Apache 2.0 @intended-inc/perimeter for OpenClaw. Decide-only: your framework executes the tool.
OpenShell →03 · At the action boundary
Your service verifies the token. Offline.
GitHub, Stripe, your cloud, your ticketing system — these are systems your agents act on, with your credentials. Govern them with the SDK or a gateway you run; verify the Authority Token against your tenant’s public keys before acting. Intended never touches them.