Skip to content

Use DoubleGate from your agent

If your agent host supports MCP, connect it to the gate directly. You do not need the Python SDK or a package manifest to make the gate's tools available.

Configure the connection

In the host's MCP server settings, provide:

  • The full HTTP MCP endpoint supplied by the gate operator.
  • The credential required by that deployment, using the host's secret facility.
  • HTTP request/response support compatible with the gate's advertised MCP version.

Host configuration formats differ; use the host's documented HTTP MCP setup rather than copying a universal-looking JSON block. This SDK supports the gate's stateless JSON response subset; SSE and sessionful MCP are not claimed by the Python client.

Verify the tools before using them

Use the host's tool discovery and confirm the server's actual catalog. The inspected Client Gate exposes these memory tasks:

Task Tool
Submit an observation doublegate.remember
Retrieve knowledge doublegate.recall
Check one artifact's state doublegate.status
Inspect its recorded history doublegate.why
Inspect pending metadata doublegate.pending

Availability and authorization are the answering gate's responsibility. Inventory is not the same as pending metadata; the inspected MCP catalog has no inventory tool.

Follow one observation through review

  1. Submit content with its source URI using doublegate.remember.
  2. Preserve the returned artifact id and inspect it with doublegate.status.
  3. Retrieve knowledge using doublegate.recall after the gate's review lifecycle.
  4. Inspect provenance and use doublegate.why when the decision needs explanation.

For authoritative recall, explicitly set include_own_pending=false. The inspected MCP tool defaults differ from the Python SDK: raw MCP may append your own unreviewed claims. Set include_provisional=false unless your application accepts provisional knowledge. Tool transport success is not an admission verdict.

Local versus remote access

The inspected Client Gate mounts MCP keylessly on loopback. Its console bearer token is checked on /api/*, not /mcp. For remote use, the operator must provide an endpoint with actual server-side authentication and authorization. Never publish the keyless local listener to the network merely because your host accepts a URL.

For application-owned memory calls instead of host tools, follow the Python quickstart. Local check manifests belong to custom check authoring, not agent connection setup.