# Use Gyblos from your code Run the local buyer and point OpenAI or Anthropic clients at it to chat with a Gyblos seller. Source: https://docs.gyblos.com/en/buyers/use-from-code The buyer is a small server on your own computer, at `http://127.0.0.1:4031`. It speaks the OpenAI and Anthropic APIs, so programs written for those services can use a Gyblos seller by changing their base URL. For each request it gets a signed free quote from the seller and streams the answer back over a direct peer-to-peer connection. ## What works today | Feature | Beta status | | -------------------------------------------------------------------- | --------------------------------------- | | Whole conversations: system message, history, new message | Yes, up to 1024 messages | | OpenAI `POST /v1/chat/completions` and Anthropic `POST /v1/messages` | Yes, whole answers and streaming | | `GET /v1/models` | Yes, the model your buyer asks for | | Free offers | Yes | | Tools (function calling), both formats, whole answers and streaming | Yes, see [Tools](#tools) | | Claude Code | Only with a seller on a larger model | | Cursor agent mode | No: it calls from Cursor's cloud | | Images, structured output, several answers | Not yet: refused with a clear 400 error | | Paid offers | Outside this guide | A seller serves each buyer up to **30 answers per minute** by default. An answer is at most **4096 tokens** unless you raise the buyer's limit (up to 32768); a longer request is served up to that limit and ends with finish reason `length`. Your prompt reaches the seller you configured, whose machine runs the model and can read it. Do not send passwords, keys or other secrets. ## Before you start - An invitation with a private **relay token**. - The **NodeId** and **owner wallet** of the seller you will use, from a source you trust (the seller prints both with `gyblos doctor`). The seller must be registered; the buyer checks its identity on Solana devnet. - Git, [Node.js](https://nodejs.org) 24.22.1 or later, and pnpm 10 (`corepack enable` provides it). In the beta the buyer runs from the source repository; the installer and the npm package contain the seller only. ## Install and configure ```sh git clone https://github.com/Gyblos/gyblos-node.git cd gyblos-node pnpm install --frozen-lockfile pnpm buyer:init ``` `buyer:init` asks for the seller's NodeId and owner wallet, shows the settings (model `qwen3-0.6b`, answers up to 4096 tokens, free quotes only) and saves them after you confirm. It sends nothing to the seller or the chain. `pnpm buyer:init --advanced` also sets the model, the answer limit and the endpoints. ## Start the buyer Type the relay token without showing it, then start the buyer in the same terminal: ```sh read -rs GYBLOS_RELAY_TOKEN && export GYBLOS_RELAY_TOKEN pnpm buyer:free ``` Leave it running. In another terminal, check that it answers: ```sh curl http://127.0.0.1:4031/v1/models ``` The list names the one model your buyer asks for; use that name as `model` in your requests. For a quick conversation in the terminal instead, run `pnpm buyer:chat`, which starts its own buyer. ## Send a conversation ```sh curl http://127.0.0.1:4031/v1/chat/completions \ -H 'content-type: application/json' \ -d '{ "model": "qwen3-0.6b", "messages": [ {"role": "system", "content": "Answer in one short sentence."}, {"role": "user", "content": "My name is Ana."}, {"role": "assistant", "content": "Nice to meet you, Ana!"}, {"role": "user", "content": "What is my name?"} ] }' ``` ### OpenAI SDK Use base URL `http://127.0.0.1:4031/v1`. The free buyer does not check the API key, but the SDK needs a non-empty one. ```js import OpenAI from 'openai'; const client = new OpenAI({ baseURL: 'http://127.0.0.1:4031/v1', apiKey: 'local', }); const stream = await client.chat.completions.create({ model: 'qwen3-0.6b', stream: true, messages: [ { role: 'system', content: 'Be brief.' }, { role: 'user', content: 'My dog is called Rocky.' }, { role: 'assistant', content: 'Got it: your dog is Rocky.' }, { role: 'user', content: "What is my dog's name?" }, ], }); for await (const chunk of stream) process.stdout.write(chunk.choices[0]?.delta?.content ?? ''); ``` ### Anthropic SDK Use base URL `http://127.0.0.1:4031` (without `/v1`). `max_tokens` is required, as with Anthropic. ```js import Anthropic from '@anthropic-ai/sdk'; const client = new Anthropic({ baseURL: 'http://127.0.0.1:4031', apiKey: 'local', }); const message = await client.messages.create({ model: 'qwen3-0.6b', max_tokens: 200, system: 'Be brief.', messages: [ { role: 'user', content: 'My dog is called Rocky.' }, { role: 'assistant', content: 'Got it: your dog is Rocky.' }, { role: 'user', content: "What is my dog's name?" }, ], }); console.log(message.content[0].text); ``` ## Tools Send `tools` as each API defines them. When the model wants a tool, the answer ends with `finish_reason: "tool_calls"` (Anthropic: `stop_reason: "tool_use"`) and carries the call. Your program runs the tool and sends the result back in the next request, with the whole conversation: ```js import OpenAI from 'openai'; const client = new OpenAI({ baseURL: 'http://127.0.0.1:4031/v1', apiKey: 'local', }); const tools = [ { type: 'function', function: { name: 'get_time', description: 'Current local time in a city', parameters: { type: 'object', properties: { city: { type: 'string' } }, required: ['city'], }, }, }, ]; const messages = [ { role: 'user', content: 'What time is it in Santiago? Use the tool.' }, ]; const first = await client.chat.completions.create({ model: 'qwen3-0.6b', messages, tools, }); const call = first.choices[0].message.tool_calls[0]; // Run the tool yourself, then send its result back. messages.push(first.choices[0].message, { role: 'tool', tool_call_id: call.id, content: '15:42', }); const second = await client.chat.completions.create({ model: 'qwen3-0.6b', messages, tools, }); console.log(second.choices[0].message.content); ``` With the Anthropic SDK, send `tools` with `input_schema`, read the `tool_use` blocks and answer with `tool_result` blocks, as with Anthropic. The buyer asks only sellers that serve tools; if none does, it answers 503 and says so. The default model, qwen3:0.6b, handles simple calls like this one. Agents that send many tools and long instructions, such as Claude Code, need a seller on a larger model through the OpenAI or Anthropic engine. ## Options and errors Options that clients send by default and that do not change the answer are accepted and ignored, such as `temperature`, `top_p`, `user`, `metadata`, `stream_options` and, in the Anthropic format, `thinking` (thinking is always off). Options that would change it are refused unless they hold their neutral value: `response_format` (`text`), `n` (1), `stop` and `stop_sequences`, and `logprobs`. Images are always refused. | Answer | Meaning | | ------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | 400 | The request uses something not supported yet; the message and `param` name it. | | 503 | No seller is available right now. With `Retry-After`, the seller failed recently or hit its per-minute limit: the OpenAI and Anthropic SDKs wait and retry on their own. | | 504 | The seller did not answer in time. | Errors use each API's own shape, so SDKs raise their usual exceptions.