Move from OpenRouter
Point an existing OpenRouter integration at Tokenlane: two values change, a few behaviors differ.
Change the connection
Set up Tokenlane first: unlock the organization, add credit and create an inference key (Quickstart, steps 1, 2 and 4). OpenRouter keys do not work here. Then swap base URL and key. Replace the entire base URL: https://openrouter.ai/api/v1 becomes https://api.tokenlane.de/v1. Changing only the host leaves /api/v1 and returns 404 not_found. Model, messages and parameters stay. The headers HTTP-Referer and X-Title may stay and are ignored.
curl
-curl 'https://openrouter.ai/api/v1/chat/completions' \
- -H "Authorization: Bearer $OPENROUTER_API_KEY" \
+curl 'https://api.tokenlane.de/v1/chat/completions' \
+ -H "Authorization: Bearer $TOKENLANE_API_KEY" \
-H "Content-Type: application/json" \
node
const client = new OpenAI({
- baseURL: "https://openrouter.ai/api/v1",
- apiKey: process.env.OPENROUTER_API_KEY,
+ baseURL: "https://api.tokenlane.de/v1",
+ apiKey: process.env.TOKENLANE_API_KEY,
});
python
client = OpenAI(
- base_url="https://openrouter.ai/api/v1",
- api_key=os.environ["OPENROUTER_API_KEY"],
+ base_url="https://api.tokenlane.de/v1",
+ api_key=os.environ["TOKENLANE_API_KEY"],
)
vercel
import { createOpenRouter } from '@openrouter/ai-sdk-provider';
-const provider = createOpenRouter({ apiKey: process.env.OPENROUTER_API_KEY });
+const provider = createOpenRouter({
+ baseURL: "https://api.tokenlane.de/v1",
+ apiKey: process.env.TOKENLANE_API_KEY,
+});
Vercel AI SDK
If you already use @openrouter/ai-sdk-provider, keep it. The connection diff above is the smallest migration; existing generateText, streamText, tool loops, structured output, reasoning and provider routing calls keep their provider.
For a new integration, @ai-sdk/openai-compatible is an alternative. Install it with ai and create the provider below; pass its model to the usual AI SDK calls. Use maxOutputTokens, temperature and other supported AI SDK settings directly. Gateway extensions such as provider, plugins and reasoning belong in providerOptions.tokenlane.
import { createOpenAICompatible } from '@ai-sdk/openai-compatible';
const tokenlane = createOpenAICompatible({
name: 'tokenlane',
baseURL: 'https://api.tokenlane.de/v1',
apiKey: process.env.TOKENLANE_API_KEY,
includeUsage: true,
});
Endpoints
Base URL https://api.tokenlane.de/v1, with the same inference key for every endpoint below. The API reference shows a runnable example for each. OpenRouter's own products (hosted agents, Auto Router, rankings) do not exist.
| Endpoint | Format | Differences |
|---|---|---|
POST /chat/completions | OpenRouter and OpenAI | Streaming, tools, structured outputs, reasoning, provider, models and plugins; details in the table below. |
[POST /responses](https://tokenlane.de/docs/gateway/post-L3Jlc3BvbnNlcw.md?lang=en) | OpenAI Responses | Stateless: previous_response_id returns 400 invalid_parameter, so send the complete conversation in input. |
[POST /messages](https://tokenlane.de/docs/gateway/post-L21lc3NhZ2Vz.md?lang=en) | Anthropic Messages | For Anthropic SDKs, set ANTHROPIC_BASE_URL to the gateway origin without /v1; the key works as x-api-key or Authorization: Bearer. |
POST /embeddings | OpenRouter and OpenAI | usage.cost in EUR. |
[POST /rerank](https://tokenlane.de/docs/gateway/post-L3JlcmFuaw.md?lang=en) | Cohere v2 style | Gateway extensions for text objects and returned documents. |
POST /audio/transcriptions, POST /audio/speech | OpenRouter | Speech defaults to pcm, not mp3. |
[POST /ocr](https://tokenlane.de/docs/gateway/post-L29jcg.md?lang=en), [POST /pii/redact](https://tokenlane.de/docs/gateway/post-L3BpaS9yZWRhY3Q.md?lang=en) | Tokenlane | No OpenRouter equivalent. |
GET /models | OpenRouter | No key needed. Prices in EUR; each model lists its routes. |
GET /key, GET /generation | OpenRouter | Amounts in EUR, also in fields named *_usd. |
What stays and what differs
| Topic | Same | Different |
|---|---|---|
| Model IDs | The vendor/model format. | Curated European catalog only. Unknown IDs, suffixes such as :free and openrouter/auto fail with 400 model_not_found. |
| Routing | provider (order, only, ignore, allow_fallbacks, require_parameters, max_price, sort) and models fallbacks. | European providers only. max_price is EUR per million tokens. sort by latency or throughput uses this gateway's measurements. Routing and fallbacks |
| Plugins | file-parser and response-healing use the familiar request shape. | Availability depends on the instance: web requires an offered search route and permission for web_search; pii-redact requires a configured detector. Without a configured detector, pii-redact returns 501 feature_unavailable. If web is not offered, it returns 400 invalid_parameter with an issue code of not_available according to the instance configuration. Do not assume that production offers them. |
| Parameters | All of OpenRouter's; unknown ones are ignored. | Ignored names appear in the x-ignored-params header. The whole body is checked before any provider call, so problems OpenRouter forwards (unknown role, broken tool schema, bad image URL) fail at once with every issue listed. |
| Errors | error.code, message, metadata; same statuses. | metadata.type is stable; request_id, issues, remedy_hint and retryable are always set. raw stays empty. Provider field paths are deliberately not imported; when no safe field is known, issues[].path keeps the $ fallback. |
| Streaming | Server-sent events ending in data: [DONE]; later failures arrive as an error event. | The last chunk before [DONE] has usage with the cost in EUR. |
| Billing | Prepaid credit, usage.cost. | EUR throughout: usage.cost, usage.currency, /models prices, max_price, key limits. Fields named *_usd in GET /v1/key hold EUR; budget objects in budgets and effective_budget explicitly carry currency: "EUR". |
| Account automation | Management keys. | The Management API on the dashboard host under /api, with its own paths and errors. Authentication |
GET /v1/byok/import-key is infrastructure for importing keys through the console. GET /v1/media/{token} is temporary provider transport. These are internal integration helpers, not endpoints for customer application code. The API reference labels them accordingly.
Deliberate endpoint and behavior differences
These OpenRouter or OpenAI endpoints are not offered and return 404 not_found:
| Not offered | Alternative |
|---|---|
GET /v1/credits | GET /v1/key, using gateway_credits for the gateway balance. |
GET /v1/providers | GET /v1/models, using the model's routes. |
GET /v1/models/{author}/{slug}/endpoints | GET /v1/models, using the model's routes. |
POST /v1/completions | POST /v1/chat/completions with a messages array. |
/api/v1 deliberately remains unsupported, not an alias for /v1.
For text-to-speech, POST /v1/audio/speech defaults to pcm, unlike OpenAI's mp3. Set response_format explicitly when the application expects a specific format.
A required feature that is permanently not configured returns 501 feature_unavailable, retryable: false, without Retry-After. Retrying the same request cannot enable it. Temporary outages remain distinct and keep their specific error type and retry guidance; not every unavailable provider or transient failure is a 501.
Before you switch traffic
- [ ] Every model ID in use is in
GET /v1/models, without suffixes. - [ ] Plugins and parameters from the table are removed or replaced.
- [ ]
max_priceand spending limits are EUR amounts. - [ ] Error handling reads
error.metadata.type; a 402 leads to a top-up, not a retry. - [ ] A normal and a streaming request succeed with a plausible
usage.costandusage.currencyEUR. - [ ] Automation built on OpenRouter management keys uses the Management API.
Next
- Privacy and compliance: data classes per route
- Errors and retries: what deserves a retry