diff --git a/README.md b/README.md
index 4687ac1..623bf6b 100644
--- a/README.md
+++ b/README.md
@@ -600,6 +600,10 @@ NADIRCLAW_COMPLEX_MODEL=claude-sonnet-4-5-20250929 \
 nadirclaw serve
 ```
 
+For a fully worked example — env vars plus a `models.local.json` pricing
+override for a hosted third-party gateway — see
+[Routing NadirClaw at a custom OpenAI-compatible gateway](docs/providers/starveri.md).
+
 ## Usage with OpenClaw
 
 [OpenClaw](https://openclaw.dev) is a personal AI assistant that bridges messaging services to AI coding agents. NadirClaw integrates as a model provider so OpenClaw's requests are automatically routed to the right model.
diff --git a/docs/providers/starveri.md b/docs/providers/starveri.md
new file mode 100644
index 0000000..7a69cd6
--- /dev/null
+++ b/docs/providers/starveri.md
@@ -0,0 +1,113 @@
+# Routing NadirClaw at a custom OpenAI-compatible gateway (Starveri example)
+
+This page is a worked example of pointing NadirClaw at a third-party
+OpenAI-compatible gateway, using [Starveri](https://api.starveri.net) as the
+concrete case. It is **not an endorsement**. Starveri is operated independently
+from OpenAI; the same recipe applies to any OpenAI-compatible endpoint
+(vLLM, LocalAI, LM Studio, OpenRouter-style gateways, etc.).
+
+No NadirClaw code change is needed for this — everything below is configuration
+you already control.
+
+## Transport: environment variables
+
+NadirClaw routes any `openai/<model>` ID through `NADIRCLAW_API_BASE` via
+LiteLLM. Set the base URL and the per-tier model IDs:
+
+```bash
+NADIRCLAW_API_BASE=https://api.starveri.net/v1 \
+NADIRCLAW_SIMPLE_MODEL=openai/gpt-5.3-codex-spark \
+NADIRCLAW_MID_MODEL=openai/gpt-5.4-mini \
+NADIRCLAW_COMPLEX_MODEL=openai/gpt-5.5 \
+nadirclaw serve --verbose
+```
+
+Use the `openai/` prefix on every model name so LiteLLM treats them as
+OpenAI-compatible. `NADIRCLAW_API_BASE` is passed to all non-Ollama,
+non-Gemini LiteLLM calls. The `NADIRCLAW_MID_MODEL` tier is optional; if you
+omit it, NadirClaw falls back to the simple model for the mid tier.
+
+## Pricing & capabilities: `models.local.json`
+
+So NadirClaw's cost tracker and savings report reflect the gateway's real
+prices, add the models to your user-managed override file at
+`~/.nadirclaw/models.local.json`. This file merges on top of the generated
+`models.json` and is never overwritten by `nadirclaw update-models`, so your
+overrides stay durable across registry refreshes.
+
+The snippet below uses the USD-per-1M-token prices published on Starveri's
+public `/models` endpoint at the time of writing. **Verify against
+<https://api.starveri.net/models> before relying on these — third-party gateway
+prices drift faster than this repo's release cadence.**
+
+```json
+{
+  "models": {
+    "openai/gpt-5.1-codex": {
+      "cost_per_m_input": 0.4166667,
+      "cost_per_m_output": 3.3333333,
+      "has_vision": false
+    },
+    "openai/gpt-5.3-codex-spark": {
+      "cost_per_m_input": 0.3333333,
+      "cost_per_m_output": 0.6666667,
+      "has_vision": false
+    },
+    "openai/gpt-5.3-codex": {
+      "cost_per_m_input": 0.1666667,
+      "cost_per_m_output": 0.3333333,
+      "has_vision": false
+    },
+    "openai/gpt-5.4": {
+      "cost_per_m_input": 0.3333333,
+      "cost_per_m_output": 1.6666667,
+      "has_vision": false
+    },
+    "openai/gpt-5.4-mini": {
+      "cost_per_m_input": 0.25,
+      "cost_per_m_output": 1.1666667,
+      "has_vision": false
+    },
+    "openai/gpt-5.5": {
+      "cost_per_m_input": 0.8333333,
+      "cost_per_m_output": 2.5,
+      "has_vision": false
+    }
+  }
+}
+```
+
+Add a `context_window` field to any entry if you want the router to enforce a
+specific input budget for that model.
+
+## Verifying the setup
+
+1. **Smoke-test each tier.** Run a request through NadirClaw with the env vars
+   set and confirm all three tiers respond. If a tier misbehaves (refusals,
+   truncation, format drift), note it — cheap tiers on third-party gateways
+   vary in quality.
+2. **Check the cost arithmetic.** Make one routed request and confirm the
+   savings report picks up the local metadata (routed cost should reflect the
+   prices above, not a fallback of $0).
+
+## Caveats
+
+- **Independent gateway.** Starveri is not OpenAI. Model IDs like `gpt-5.5`
+  are the gateway's labels for whatever it serves behind them.
+- **Prices drift; `/models` is the source of truth.** The numbers above are a
+  point-in-time snapshot. Re-pull <https://api.starveri.net/models> periodically
+  and update your `models.local.json`.
+- **Cached-input and per-tool fees are not modeled.** NadirClaw's metadata
+  schema currently has only `cost_per_m_input` and `cost_per_m_output`.
+  Gateways that also bill cached-input tokens or per-tool-call surcharges
+  (Starveri publishes both) will cost slightly more than NadirClaw's tracker
+  reports. Treat the savings figure as a close estimate, not an invoice.
+- **Not a permanent free provider.** Starveri offers a limited text-only demo,
+  but paid usage is prepaid credits. Don't configure it expecting free routing.
+
+## See also
+
+- [Usage with Custom OpenAI-Compatible Endpoints](../../README.md#usage-with-custom-openai-compatible-endpoints)
+  — the general recipe this page specializes.
+- [Multi-provider routing](../multi-provider-routing.md) — routing across more
+  than one vendor and the failure modes that introduces.