Self-hosted, coding-only completion API at api.misar.io/coder/offline/* — REST, SSE and WebSocket over one schema.

A separate surface from the cloud gateway

This is the offline coding API, hosted at https://api.misar.io/coder/offline. It runs MisarCoder's own self-hosted models (Ollama, CPU) — no cloud AI provider — and is deliberately coding-only. It is distinct from the cloud MoE gateway at api.misar.io/coder documented on the other pages in this section.

Base URL: https://api.misar.io/coder/offline Auth: Authorization: Bearer <key> on every call. Models: coder (qwen2.5-coder 7B, default) · coder-fast (qwen2.5-coder 3B, faster on CPU).

Endpoints

Transport	Endpoint	Purpose
REST	`POST /v1/code/completions`	buffered code completion / chat
SSE	`POST /v1/code/completions` with `"stream": true`	`text/event-stream` token stream
WebSocket	`/v1/code/session`	interactive, cancellable coding session
REST	`GET /v1/code/models`	model catalogue
REST	`GET /health`	liveness

REST — code completion

POST/v1/code/completions

Generate a code completion. Provide either prompt (single-shot) or messages (multi-turn). Set stream: true for SSE; otherwise a single JSON object is returned.

Request body

promptstringbody

Single-shot instruction. Provide this or messages.

messagesArray<{role, content}>body

Multi-turn conversation. Provide this or prompt. role is system | user | assistant.

modelstringbody

coder (default) or coder-fast. Unknown values fall back to coder.

systemstringbody

Override the default coding system prompt.

temperaturenumberbody

Sampling temperature. Defaults to the server's coding default.

max_tokensintegerbody

Maximum tokens to generate.

streambooleanbody

When true, returns a text/event-stream instead of a JSON body.

curl -X POST https://api.misar.io/coder/offline/v1/code/completions \
  -H "Authorization: Bearer $MISARCODER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"coder","prompt":"Write a Python is_prime(n) function. Return only code."}'

{
  "id": "cmpl_1a2b3c…",
  "object": "code.completion",
  "created": 1750000000,
  "model": "coder",
  "resolved_model": "qwen2.5-coder:7b-instruct-q4_K_M",
  "choices": [
    { "index": 0, "message": { "role": "assistant", "content": "def is_prime(n): …" }, "finish_reason": "stop" }
  ],
  "usage": { "prompt_tokens": 24, "completion_tokens": 96, "total_tokens": 120 }
}

An invalid or missing bearer key returns 401. A request with neither prompt nor messages returns 400.

SSE — token stream

Send the same body with "stream": true. The response is text/event-stream. Each delta is a data: line; a terminal event carries done: true + usage, followed by data: [DONE].

data: {"id":"cmpl_…","model":"coder","delta":"def "}

data: {"id":"cmpl_…","model":"coder","delta":"is_prime"}

data: {"id":"cmpl_…","model":"coder","done":true,"finish_reason":"stop","usage":{"prompt_tokens":24,"completion_tokens":96,"total_tokens":120}}

data: [DONE]

WebSocket — interactive session

Connect to wss://api.misar.io/coder/offline/v1/code/session. Authenticate with a ?key=<key> query param, an Authorization header, or a first {"type":"auth","token":"…"} frame.

Client → server

Message	Fields
`{"type":"prompt", …}`	`prompt` \| `messages`, plus optional `model`, `system`, `temperature`, `max_tokens`, `id`
`{"type":"cancel"}`	abort the in-flight turn
`{"type":"ping"}`	heartbeat → server replies `{"type":"pong"}`

Server → client

Message	Meaning
`{"type":"start","id","model","resolved_model"}`	turn began
`{"type":"delta","id","text"}`	streamed token(s)
`{"type":"done","id","finish_reason","usage"}`	turn complete
`{"type":"error","message"}`	failure

SDKs

Same params, same response shape

Both SDK clients mirror this contract exactly across REST, SSE and WebSocket.

import { CoderClient } from "@misar/agent-sdk";

const coder = new CoderClient({
  baseUrl: "https://api.misar.io/coder/offline",
  apiKey: process.env.MISARCODER_API_KEY,
});

// REST
const { content } = await coder.complete({ prompt: "Write is_prime(n) in Python." });

// SSE
await coder.stream({ prompt: "reverse a string in python", model: "coder-fast" },
  (t) => process.stdout.write(t));

// WebSocket
const session = await coder.session();
await session.prompt({ prompt: "quicksort in Go" }, { onDelta: (t) => process.stdout.write(t) });
session.close();

from misarcoder import CoderClient

coder = CoderClient(api_key="csk_…")            # base_url defaults to api.misar.io/coder/offline
r = coder.complete(prompt="Write is_prime(n) in Python.")
print(r["choices"][0]["message"]["content"])

for chunk in coder.stream(prompt="reverse a string", model="coder-fast"):
    if chunk.get("delta"):
        print(chunk["delta"], end="")

Latency (CPU inference)

The offline models run on a CPU-only backend. Expect roughly ~1.4 tok/s (coder, 7B) and ~2.4 tok/s (coder-fast, 3B) when warm; the first request to a cold model adds ~25–60 s of load time. Prefer coder-fast for interactive use.

Offline Coding API

Endpoints

REST — code completion

SSE — token stream

WebSocket — interactive session

SDKs