Cohere Free code examples in C#, Python, Java and JavaScript

Note the environment variable is CO_API_KEY, not COHERE_API_KEY — a small thing that costs ten minutes the first time.

pip install cohere

export CO_API_KEY="..."      # macOS / Linux
setx CO_API_KEY "..."        # Windows, then open a new shell

ClientV2, not Client. The v1 client takes a single message string; v2 takes a messages array like everything else on this page. Response text is nested one level deeper than OpenAI's.

import cohere

client = cohere.ClientV2()      # reads CO_API_KEY

response = client.chat(
    model="command-a-03-2025",
    messages=[
        {"role": "system", "content": "You are a concise technical assistant."},
        {"role": "user", "content": "Explain a database index in two sentences."},
    ],
)

# content is a list of typed blocks, not a string.
print(response.message.content[0].text)
print(response.usage.tokens)

Events are typed by a string tag. Anything other than content-delta — citations, tool calls, the end marker — will crash a loop that assumes text is always there.

stream = client.chat_stream(
    model="command-a-03-2025",
    messages=[{"role": "user", "content": "Write a haiku about indexes."}],
)

for event in stream:
    if event.type == "content-delta":
        print(event.delta.message.content.text, end="", flush=True)

print()

This, not chat, is why most teams add Cohere. Vector search gets you twenty roughly-relevant chunks; rerank orders them by actual relevance to the question and lets you send three instead. Less context, lower cost, better answers — and it works on top of whatever retrieval you already have.

import cohere

client = cohere.ClientV2()

# Whatever your vector store returned — over-fetch deliberately.
candidates = [
    "Annual leave accrues at 1.75 days per month of service.",
    "The office is closed on national holidays.",
    "Notice period is 60 days for all permanent staff.",
    "Expense claims must be filed within 30 days.",
]

ranked = client.rerank(
    model="rerank-v3.5",
    query="How much notice do I have to give?",
    documents=candidates,
    top_n=2,
)

for result in ranked.results:
    print(f"{result.relevance_score:.3f}  {candidates[result.index]}")

# Send only the top_n to the model. Over-fetch wide, rerank, send narrow.

Cohere ships an official Java SDK (com.cohere:cohere-java), but its README still shows the v1 chat shape while Python is on v2. These samples use the v2 HTTP API so the page teaches one API — swap in the SDK once you have checked which version yours exposes.

<!-- Nothing required: java.net.http is in the JDK since Java 11. -->
<!-- Official SDK, if you prefer it and have checked its version:  -->
<dependency>
    <groupId>com.cohere</groupId>
    <artifactId>cohere-java</artifactId>
    <version>1.6.0</version>
</dependency>

java.net.http.HttpClient, no dependency. Response text is nested: message.content[0].text.

import com.fasterxml.jackson.databind.JsonNode;
import com.fasterxml.jackson.databind.ObjectMapper;

import java.net.URI;
import java.net.http.HttpClient;
import java.net.http.HttpRequest;
import java.net.http.HttpResponse;
import java.util.List;
import java.util.Map;

ObjectMapper mapper = new ObjectMapper();
HttpClient http = HttpClient.newHttpClient();

String payload = mapper.writeValueAsString(Map.of(
    "model", "command-a-03-2025",
    "messages", List.of(
        Map.of("role", "system", "content", "You are a concise technical assistant."),
        Map.of("role", "user", "content", "Explain a database index in two sentences."))));

HttpRequest request = HttpRequest.newBuilder()
    .uri(URI.create("https://api.cohere.com/v2/chat"))
    .header("Authorization", "Bearer " + System.getenv("CO_API_KEY"))
    .header("Content-Type", "application/json")
    .POST(HttpRequest.BodyPublishers.ofString(payload))
    .build();

HttpResponse<String> response = http.send(
    request, HttpResponse.BodyHandlers.ofString());

if (response.statusCode() != 200) {
    System.err.println("Cohere " + response.statusCode() + ": " + response.body());
    return;
}

JsonNode body = mapper.readTree(response.body());
System.out.println(body.get("message").get("content").get(0).get("text").asText());

The reason most teams add Cohere. Over-fetch from your vector store, rerank, send only the top few to the model.

import com.fasterxml.jackson.databind.JsonNode;

List<String> candidates = List.of(
    "Annual leave accrues at 1.75 days per month of service.",
    "The office is closed on national holidays.",
    "Notice period is 60 days for all permanent staff.",
    "Expense claims must be filed within 30 days.");

String payload = mapper.writeValueAsString(Map.of(
    "model", "rerank-v3.5",
    "query", "How much notice do I have to give?",
    "documents", candidates,
    "top_n", 2));

HttpRequest request = HttpRequest.newBuilder()
    .uri(URI.create("https://api.cohere.com/v2/rerank"))
    .header("Authorization", "Bearer " + System.getenv("CO_API_KEY"))
    .header("Content-Type", "application/json")
    .POST(HttpRequest.BodyPublishers.ofString(payload))
    .build();

JsonNode body = mapper.readTree(
    http.send(request, HttpResponse.BodyHandlers.ofString()).body());

for (JsonNode result : body.get("results")) {
    int index = result.get("index").asInt();
    double score = result.get("relevance_score").asDouble();

    System.out.printf("%.3f  %s%n", score, candidates.get(index));
}

Three things that differ from every other provider on this page.

// 1. THE ENV VAR IS CO_API_KEY, not COHERE_API_KEY.

// 2. MODEL IDS MOVED. command-r-plus-08-2024 appears in most tutorials
//    and is two generations old; command-a-03-2025 is current.

// 3. RESPONSE TEXT IS NESTED — message.content[0].text. content is an
//    array of typed blocks so tool calls and citations sit beside the
//    text, which is why there is no single content string.

// Rerank is the endpoint worth the integration. Chat is competent;
// rerank improves a retrieval pipeline you already have.

Cohere ships an official SDK (cohere-ai). These samples use the v2 HTTP API instead, because the SDK's chat surface differs by version and the HTTP shape is the one thing that is the same everywhere — swap the SDK in once you have checked which version yours exposes.

# Nothing to install — fetch is built into Node 18+.
# Official SDK, if you prefer it:  npm install cohere-ai

export CO_API_KEY="..."

Bearer auth, and the response nests text one level deeper than OpenAI: message.content is an array of typed blocks.

const response = await fetch("https://api.cohere.com/v2/chat", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.CO_API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    model: "command-a-03-2025",
    messages: [
      { role: "system", content: "You are a concise technical assistant." },
      { role: "user", content: "Explain a database index in two sentences." },
    ],
  }),
});

if (!response.ok) {
  throw new Error(`Cohere ${response.status}: ${await response.text()}`);
}

const body = await response.json();
console.log(body.message.content[0].text);

This, not chat, is why most teams add Cohere. Over-fetch from your vector store, rerank, then send only the top few to the model: less context, lower cost, better answers.

const candidates = [
  "Annual leave accrues at 1.75 days per month of service.",
  "The office is closed on national holidays.",
  "Notice period is 60 days for all permanent staff.",
  "Expense claims must be filed within 30 days.",
];

const response = await fetch("https://api.cohere.com/v2/rerank", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.CO_API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    model: "rerank-v3.5",
    query: "How much notice do I have to give?",
    documents: candidates,
    top_n: 2,
  }),
});

const { results } = await response.json();

for (const result of results) {
  console.log(result.relevance_score.toFixed(3), candidates[result.index]);
}

Three things that differ from every other provider on this page.

// 1. THE ENV VAR IS CO_API_KEY, not COHERE_API_KEY.

// 2. MODEL IDS MOVED. command-r-plus-08-2024 appears in most tutorials
//    and is two generations old; command-a-03-2025 is current. Keep it
//    in config.
const MODEL = process.env.COHERE_MODEL ?? "command-a-03-2025";

// 3. RESPONSE TEXT IS NESTED. body.message.content[0].text, not
//    choices[0].message.content — content is an array of typed blocks
//    so that tool calls and citations can sit beside the text.