Most AI gateways are built for applications that call models. Agents need a slightly different contract - constrained tool access and clear attribution. We just shipped Phase 1 of agent-native support on Leanroute.
1. Discover allowlisted MCP servers
GET /v1/mcp/servers
Authorization: Bearer gw_live_...
Returns only the MCP servers on your organization’s allowlist. Agents no longer need hardcoded URLs, and they can’t see servers that aren’t approved.
2. Attribute work to agent runs
Send optional headers:
-
X-Agent-Name— which agent -
X-Agent-Session-Id- which run
Leanroute then groups requests and reports:
- Cost per run
- Models used
- Tools called (MCP tools executed at the gateway, plus tools the model requested)
You’ll find this in:
- Dashboard → Agent runs
- GET /platform/v1/runs
- MCP tools: list_runs, get_run
Tool arguments and results are not stored.
Design constraints we kept
- Gateway remains stateless
- Agent memory/state stays on the agent side
- No agent builder, no global registry
- Org MCP allowlist is still the enforcement boundary
If you’re wiring agents through an AI gateway and want enforcement plus visibility without adopting a full agent platform, this is the layer we think belongs at the gateway.
Visit our Docs for more details
Top comments (0)