Getting started — first cost report in 30 seconds

Format your log as JSONL or JSON array

Each line (JSONL) or array entry is one LLM call. Required fields: provider, model, prompt_tokens, completion_tokens. Optional: timestamp, user_id, cache_read_tokens, cache_write_tokens, error.

{"provider":"anthropic","model":"claude-opus-4-7","prompt_tokens":500,"completion_tokens":200,"timestamp":"2026-05-15T10:00:00Z","user_id":"alice"}
{"provider":"openai","model":"gpt-5-mini","prompt_tokens":1000,"completion_tokens":500}
{"provider":"google","model":"gemini-2.5-flash","prompt_tokens":2000,"completion_tokens":300}

Don't send prompt or completion text — the Actor doesn't need it and will ignore those fields.

Open the Actor + paste your log

Go to apify.com/charcoal_jam/tokenmark-llm-cost-analyzer. Set Operation to analyze. Paste your JSONL into the JSONL log field (or pass entries as a JSON array via the API).

Click "Start" — get a report in seconds

The Actor processes your log (typically <5s for 1k entries). The dataset view "Summary" shows the structured output as a table. Or grab the JSON via run-sync-get-dataset-items and pipe it anywhere.

curl -X POST "https://api.apify.com/v2/acts/charcoal_jam~tokenmark-llm-cost-analyzer/run-sync-get-dataset-items?token=$APIFY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"operation":"analyze","jsonl":"..."}'

Install + wrap your client

npm install tokenmark

// In your code (Anthropic example):
import Anthropic from "@anthropic-ai/sdk";
import { wrapAnthropic } from "tokenmark";

const client = wrapAnthropic(new Anthropic());
// Use client like normal — every call is logged to ./tokenmark.jsonl

Or for OpenAI: import { wrapOpenAI } from "tokenmark"; wrapOpenAI(new OpenAI()).

Run your app, accumulate the log

tokenmark logs every wrapped call to ./tokenmark.jsonl (or wherever TOKENMARK_SINK points). Run your app normally — no other code changes.

Query the log

npx tokenmark report --since 7d --route-recommendations --json

Outputs the same shape as the Apify Actor's analyze response — summary, by_day, by_model, by_user, top_costly_calls, recommendations.

For autonomous agents: expose the log via the MCP server:

// Claude Desktop / Cursor / Cline config:
{
  "mcpServers": {
    "tokenmark": {
      "command": "npx",
      "args": ["-y", "tokenmark-mcp"]
    }
  }
}