Claude Desktop via RelayVano: API Relay Setup for Document Work

The pairing mistake that breaks document workflows

Document-heavy work punishes guesswork. When you push a 60-page contract, a research packet or a stack of meeting notes through an AI client, the small mistakes turn into large ones: the wrong client for the model family, the wrong endpoint, and a bill you cannot explain when the month closes.

The most common mistake is assuming that one universal endpoint drives every model from every tool. RelayVano is an independent relay made for developers worldwide, and it is deliberately explicit about how pairing works. Claude models work only in the official Claude Code and Claude Desktop clients, version 2.1.285 or newer. Gemini models use the native Gemini protocol only, in Gemini CLI. Codex, Cursor, Cherry Studio, Cline, OpenCode, OpenAI SDKs and cURL pair only with GPT, DeepSeek and the other non-Claude, non-Gemini models.

That sounds narrow until you have lost an afternoon to a client that quietly refuses a model. Knowing the rule before you start is faster than learning it through error messages.

Here is the good news: one API key reaches every model through https://api.relayvano.com, with three routes: OpenAI-compatible /v1, Anthropic /v1/messages and native Gemini /v1beta. So the setup is one key and one base address, and the only thing you have to get right is which client holds which model.

Setting up Claude Desktop on RelayVano

For document-heavy work with Claude, the path is short. First, create your RelayVano account and copy the API key from the dashboard. Next, open Claude Desktop and confirm you are on version 2.1.285 or newer, because that is the version range RelayVano supports for Claude models. Then point the Anthropic-compatible base address in the client to https://api.relayvano.com, using the /v1/messages route for Claude requests.

After that, choose the Claude model that matches the job and send a small test document before you upload the whole archive. A single-page test tells you in seconds whether the key, the base address and the model family all line up. If the reply comes back, you can move to a real batch with confidence.

A few habits keep this tidy. Keep one key in the client rather than pasting a fresh one into every project, since a single key is all it takes. Test with a document you know well, so you can judge the output quality and the cost together. And do not try to force a Claude model into Codex, Cursor, Cherry Studio, Cline, OpenCode, an SDK or cURL: those pair only with GPT, DeepSeek and the other non-Claude, non-Gemini models, and mixing them up is the classic wasted afternoon.

Because RelayVano serves developers worldwide, the site, the dashboard and the documentation all come in six languages: 中文, English, Español, हिन्दी, Português, Français. Prices display in yuan on the Chinese pages and in approximate dollars on the others, so a teammate in another country reads the same instructions and the same numbers you do.

Reading Claude fares before you send the batch

The second big mistake is sending a thousand document requests without checking what a request actually costs. RelayVano bills on a prepaid balance, pay per token, with no monthly fee and a balance that never expires. Prompt tokens, reply tokens, cache stores and cache hits are priced separately, so the shape of your document workflow matters as much as its size.

Consider Claude Sonnet 5.5 at an approximate $1.25 prompt, $6.25 reply and $0.0889 cache hit per million tokens, against an official list price of $2, $10 and $0.1. Claude Opus 5.5 runs about $1.65 prompt, $7.82 reply and $0.075 cache hit, versus an official $4, $20 and $0.2. For lighter classification and extraction passes, Claude Haiku 4.5 sits at about $0.3958 prompt, $1.90 reply and $0.0319 cache hit, compared with an official $1, $5 and $0.1. These fares are approximate US dollars per million tokens at roughly 1 USD ≈ ¥7.2, and the actual charge is made in yuan.

Here is an example, not a quote. Say you send 200 requests a day, each carrying 1,000 prompt tokens and returning 500 reply tokens on Claude Sonnet 5.5. That is 200,000 prompt tokens, or 0.2 million, at $1.25, which comes to $0.25. The replies total 100,000 tokens, or 0.1 million, at $6.25, which comes to $0.625. Together that is about $0.875 for the day, or roughly $26.25 across 30 days of that same pattern. Work the arithmetic with the model you actually plan to use, because a heavier reasoning model changes the reply column considerably.

The lesson is simple: model selection is a cost decision. Routing routine extraction to a lighter Claude model and reserving the heavier one for judgment-heavy documents is where the pay-as-you-go model starts to feel like good value.

Cache pricing, usage visibility and interrupted streams

The third mistake is assuming that a disconnected session is free money. If a caller disconnects mid-stream, the request is still finished and billed on the tokens really used. The work happened, so the cost is real. Planning for long document responses and stable connections is more predictable than gambling on a dropped connection.

The fourth mistake is flying blind on spending. On RelayVano, every request shows up with its charge on the usage page. When a batch of document runs costs more than expected, you can see which requests drove it instead of guessing from a monthly total. Combined with cache stores and cache hits being priced separately, that visibility lets you compare a run that reuses cached context against one that rebuilds it from scratch, and pick the cheaper path for the next batch.

Because the balance never expires, you are not pushed into a vendor's subscription to keep access. You top up what you need, spend it down at your own pace, and read the charges line by line while the work is happening.

Top-ups, refunds and support in six languages

Money movement should be as predictable as the API. Top-ups go through Stripe checkout: cards work everywhere, plus WeChat Pay where your account supports it. If a top-up is larger than you ended up needing, unspent top-ups can be refunded within 9 days. That window makes it practical to fund a heavier document month and still recover the difference if the project changes shape.

Referrals work in your favour too. You earn 16% of every top-up by people who sign up through your link, and 4% one level further. Earnings can be withdrawn from $5.56 or moved straight into the balance to cover usage, which turns a recommendation into working credit rather than a side note.

When something needs explaining, logged-in members write to us under Messages in the dashboard. Support lives in the same place as the usage page and the top-up button, and the interface around it is available in 中文, English, Español, हिन्दी, Português and Français. A team spread across time zones still reads one consistent workflow.

Starting your first document run

New accounts start with $0.40 of credit, which is enough to send a test document, confirm the pairing and see the first charge appear on the usage page. From there, the loop is the same every time: top up, send, read the charge, adjust the model or the context strategy, and repeat.

RelayVano gives you one API key, a base address at https://api.relayvano.com, clear client-to-model pairing, token-level pricing you can read before you commit, and a prepaid balance that never expires. For document-heavy work in Claude Desktop, that combination removes the two failure modes that cost the most time: the wrong setup and the unexplained bill.

Create your account, add your first top-up, and point Claude Desktop at RelayVano. You can begin with the $0.40 starter credit and see for yourself how the pieces fit together.