diff --git a/docs/E2E_FREE_MCP_INFERENCE_CHECKLIST.md b/docs/E2E_FREE_MCP_INFERENCE_CHECKLIST.md
new file mode 100644
index 0000000..0f4ed91
--- /dev/null
+++ b/docs/E2E_FREE_MCP_INFERENCE_CHECKLIST.md
@@ -0,0 +1,10 @@
+# E2E Checklist — agent-cli (Phase 4)
+
+See web-auth `docs/E2E_FREE_MCP_INFERENCE_CHECKLIST.md` for full stack flows.
+
+## agent-cli specific
+
+- [ ] MCP connect without subscription ID still uploads metering events
+- [ ] `openrouter_chat` requires inference plan at gateway (not agent-cli local gate)
+- [ ] Builder fee regression: `tests/test_builder_fee.py` unchanged pass
+- [ ] Exhibit A harness documents free MCP vs inference paths (`docs/FREE_MCP_METERING.md`)
diff --git a/docs/FREE_MCP_METERING.md b/docs/FREE_MCP_METERING.md
new file mode 100644
index 0000000..542f323
--- /dev/null
+++ b/docs/FREE_MCP_METERING.md
@@ -0,0 +1,21 @@
+# Free MCP metering (no inference subscription)
+
+As of the 2026-07-07 product split, **MCP trading tools are free** and monetized via Hyperliquid builder code. An inference subscription is optional and only required for Nunchi/OpenRouter-backed `openrouter_chat`.
+
+## Gateway behavior
+
+- `mcp-gateway` calls web-auth `/api/metering/status` **only when**:
+  - the connect payload includes `subscription_id` (inference subscriber), or
+  - the tool is `openrouter_chat`.
+- Free MCP tool calls (`trade`, `funding_hedge_execute`, market reads, etc.) proceed without subscription quota checks.
+
+## agent-cli expectations
+
+- `POST /api/mcp/connect` may omit `planId` / subscription fields for free MCP.
+- When no inference subscription is active, web-auth mints gateway tokens with `plan_id: free-mcp` and no `subscription_id`.
+- Usage upload (`POST /api/metering/usage`) still records control-plane and tool events for observability; it does not gate free MCP trading.
+- BYO inference (user-supplied model keys) stays outside Nunchi metering unless the user purchases an inference plan.
+
+## When to pass subscription context
+
+Pass `subscription_id` and inference `plan_id` from web-auth connect only when the user has an active inference subscription. agent-cli should not fabricate subscription IDs for free MCP sessions.