>The Compute API (Developer Demand)

The Compute API is the developer door to the same network: the same inference, the same nodes, the same Proof-of-Execution, the same $PRLX credit balance, reached programmatically instead of through a chat box. It exists so that a script, an OpenAI-compatible tool, or an AI agent can run inference on the network without a human in the loop. It is served at api.parallelix.io and authenticated by API keys.

The API presents one billing rail behind three request shapes:

1. OpenAI-compatible endpoint (POST /v1/chat/completions). Any existing tool that speaks the OpenAI chat-completions format (an editor plugin, an orchestration framework, a script) points at ParalleliX by changing one base URL. The request and response envelopes match the OpenAI shape, with an extra parallelix object carrying the serving node id and the Proof-of-Execution hash.
2. Parallel batch endpoint (POST /v1/batch). Submit an array of prompts; the coordinator enqueues each as an independent inference request and fans them out across the online nodes simultaneously. This is the parallel-native shape an OpenAI-compatible call cannot express: one submission, N sub-tasks executing at once, each returning its own result and PoE. Poll GET /v1/batch/{id} for per-item status.
3. MCP connector (parallelix-mcp). A Model Context Protocol server (published to npm) that wraps the two endpoints as tools an MCP client (Claude Code, Claude Desktop) can call. The intended pattern: a frontier agent orchestrates and reasons; the distributed open-source fleet executes the cheap, embarrassingly-parallel parts (bulk classify, extract, summarize, translate) through a single parallel_map tool call.