Hermes Agent Token Counting Methods | Generated by AI

Home 2026.05

Good question. Since you’re using MiMo via the Anthropic-compatible endpoint with Claude Code / agent tooling, the token counts come back in the API response’s usage field — same as standard Anthropic API.Question: In Hermes Agent, how do you see input and output token counts?

Answer:

Several ways, from built-in to DIY:

1. Built-in: /analytics page

Hermes has an Analytics page at /analytics that displays a Daily Token Usage chart with input and output token counts. Navigate there in the Hermes web UI. Note: there’s a known bug where input is reported far smaller than output — input/output counts may appear swapped or undercounted in the insights report.

2. Session files: ~/.hermes/sessions/

Hermes stores request dumps in ~/.hermes/sessions/. You can inspect them directly:

# List sessions
ls ~/.hermes/sessions/

# Grep for token usage across all sessions
grep -r "input_tokens\|output_tokens" ~/.hermes/sessions/ | tail -50

# Pretty-print a specific session
cat ~/.hermes/sessions/<session_id>.json | python3 -m json.tool | grep -A5 "usage"

The usage object in each API response follows the standard Anthropic shape:

{
  "input_tokens": 14200,
  "output_tokens": 843,
  "cache_read_input_tokens": 11500,
  "cache_creation_input_tokens": 0
}

3. What the token counts actually mean for MiMo Credits

Given the MiMo pricing table from the doc, you can calculate Credit burn yourself. For mimo-v2.5:

def credits_used(input_tokens, output_tokens, cache_hit_tokens=0):
    cache_miss = input_tokens - cache_hit_tokens
    credits = (
        cache_hit_tokens   * 2      +   # 2 Credits/token cache hit
        cache_miss         * 100    +   # 100 Credits/token cache miss
        output_tokens      * 200        # 200 Credits/token output
    )
    return credits

# Example: typical Hermes call
print(credits_used(14000, 800, cache_hit_tokens=11500))
# cache hit: 11500*2=23000, miss: 2500*100=250000, out: 800*200=160000
# Total: 433,000 Credits per call

With Lite plan (4.1B Credits), that’s ~9,470 calls before quota exhausted. Fits the “~200 medium-complex tasks” estimate from the official docs (each task = multiple calls).

4. Reality check on Hermes overhead

A typical Hermes API call’s input is dominated by: tool definitions (~8K tokens), system prompt (~2–3K tokens), SOUL/MEMORY/USER.md (~1–5K tokens), and skills if loaded (~0–2K tokens). For a brand-new chat with no skills active, the input cost is roughly 12K–14K tokens before you’ve sent your first message.

So for MiMo Credit math: assume ~13K input tokens mostly as cache miss on the first call, then mostly cache hits on subsequent turns within the same session.

References:

Back Donate