How do you use the Plugsky chat completions API?

Key facts

Endpoint	POST https://plugsky.com/v1/chat/completions
Request shape	model, messages and optional temperature, top_p, n, stop, max_tokens, penalties, tools, response_format
Response shape	Same id, object, created, model, choices, finish_reason and usage fields as OpenAI
Streaming	stream: true returns the OpenAI SSE delta format
Tools	tools and tool_choice accept OpenAI-format function definitions
JSON mode	response_format: {type: json_object}
Routing	Set model to plugsky-fusion to route per request
Product status	Live

TL;DR

• Same request and response JSON as OpenAI — change the base URL and model.
• Streaming, function calling, JSON mode and vision all use OpenAI shapes.
• Model is required: pick a Plugsky model or plugsky-fusion for routing.
• Errors use the OpenAI error schema, so existing retry logic keeps working.
• Start free with plugsky-micro or plugsky-lite, then move up tiers.

How it works, step by step

1. Generate an API key in the Plugsky dashboard.
2. Point your OpenAI client base_url at https://plugsky.com/v1.
3. Send model and messages — the minimum required fields.
4. Add tools or response_format when you need function calling or JSON.
5. Turn on stream: true for token-by-token SSE responses.
6. Watch usage in the response body or the dashboard usage view.

Try it yourself

Open the OpenAI-compatible API tester →

The endpoint and request shape

The endpoint is POST https://plugsky.com/v1/chat/completions. Only model and messages are required. Messages carry one of four roles — system, user, assistant or tool — and optional fields cover sampling (temperature, top_p, seed), length (max_tokens), repetition (presence_penalty, frequency_penalty) and stopping (stop).

For generation count use n; for safety attribution use user. Because the schema matches OpenAI, prompts, SDK calls and validation code port without changes.

Streaming, tools and JSON mode

Set stream: true to receive server-sent events with the same delta structure as OpenAI. Function calling uses a tools array of JSON-Schema definitions plus tool_choice set to none, auto or a named function; the model returns structured tool calls you execute and send back as a tool-role message. JSON mode is a single field: response_format: {type: json_object}.

• Keep streaming and tools separate per request for simpler error handling.
• Validate JSON-mode output against your schema before writing it to storage.
• Use seed for best-effort determinism in tests, not as a correctness guarantee.

Vision and model choice

Vision models accept content as an array of text and image_url parts, matching OpenAI's vision format. Model choice is explicit: the catalogue spans 30+ models from free chat models to frontier reasoning, plus plugsky-fusion if you want the platform to pick per request. You can also set a fixed model per API key when you want a predictable default.

Errors, limits and compatibility notes

Errors use the OpenAI schema with message, type, code and param fields, and rate limits return 429 with a Retry-After header. Request bodies are capped at 16 MB, and context overflows return a 400 that includes exact token counts.

Honest caveats: legacy /v1/completions and the stateful /v1/responses endpoint are listed as coming soon, so new builds should target chat completions. Audio, images and fine-tuning are also roadmap items — check the docs before planning those workloads.

Honest comparison

Capability	Plugsky chat completions	Typical OpenAI-compatible provider	Building a compatibility layer in-house
Request/response shape	Identical to OpenAI	Usually compatible	You maintain parity
Streaming	OpenAI SSE deltas	Common	You implement SSE
Function calling	OpenAI tools format	Varies by provider	You adapt per model
Model choice	30+ models behind one endpoint	Varies	You integrate each
Error schema	OpenAI error shape	Varies	You normalise errors
Migration effort	Change base_url and model	Often a rewrite	Months of work