Cohere - Zespan

Available for: Python and TypeScript.

Wrap your Cohere ClientV2 with wrapCohere() (TypeScript) to trace chat() and chatStream() calls — or patch the cohere module with patch_cohere() (Python).

Installation

npm install @zespan/sdk cohere-ai

pip install zespan cohere

Setup

import { CohereClientV2 } from "cohere-ai";
import { wrapCohere } from "@zespan/sdk";
import { zespan } from "@zespan/sdk";

zespan.init({ apiKey: process.env.ZESPAN_API_KEY! });

const cohere = wrapCohere(new CohereClientV2({ token: process.env.COHERE_API_KEY! }));

import os
import zespan

zespan.init(api_key=os.environ["ZESPAN_API_KEY"])
zespan.patch_cohere()

import cohere  # import after patching
client = cohere.ClientV2(api_key=os.environ["COHERE_API_KEY"])

In TypeScript, wrapCohere() returns a wrapped client instance. In Python, patch_cohere() patches ClientV2.chat and ClientV2.chat_stream at the class level, so every client constructed afterward is traced automatically.

Example

const response = await cohere.chat({
  model: "command-r-plus",
  messages: [{ role: "user", content: "Explain embeddings briefly." }],
});

response = client.chat(
    model="command-r-plus",
    messages=[{"role": "user", "content": "Explain embeddings briefly."}],
)

chatStream() (TypeScript) and chat_stream() (Python) are traced end-to-end in both languages, including time-to-first-token — parsed from Cohere’s content-delta, tool-call-start/tool-call-delta, and message-end stream events.

What gets captured

Field	Details
Model	`command-r-plus`, `command-r`, `command-light`, etc.
Input tokens	From `usage.billed_units.input_tokens` (falls back to `usage.tokens`)
Output tokens	From `usage.billed_units.output_tokens` (falls back to `usage.tokens`)
Cost	Calculated from token counts and Cohere pricing
Latency	Total request duration
Finish reason	`COMPLETE`, `MAX_TOKENS`, `TOOL_CALL`, `ERROR`
Tool calls	Tool name and parsed arguments, both languages

Cohere’s response text lives in message.content, a list of typed blocks ({type: "text", text: "..."}) rather than a plain string — the wrapper flattens this automatically before storing prompt/completion text.