Mycroft

Claude Code spend, read from Grafana or Datadog

The cheaper model isn't always cheaper.

Mycroft reads the Claude Code metrics you already send to Grafana or Datadog, prices every token with the prompt cache counted, and tells you which changes would actually pay.

Get a free report Read a sample finding 
MYCROFT
Anthropic usage - 1-28 August 2026
==============================================

WHAT IT COST
  $18,412.60 at list prices

  by request kind
     $12,907.41   70.1%  main
      $3,871.22   21.0%  auxiliary
      $1,633.97    8.9%  subagent

----------------------------------------------
FINDING 2  21.0% of spend is the agent's own
  housekeeping, on claude-opus-5
----------------------------------------------
  No cache on claude-opus-5 for this traffic,
  so moving it to claude-haiku-4-5 costs
  nothing to switch.

  At most $3,096.98 over this period
    It is a ceiling, not a saving.
Sample report for an invented team. The arithmetic is real.

Switching models throws away the prompt cache.

On a typical agent step, the first request after moving from Sonnet 4.6 to Haiku 4.5 costs 3.9 times what staying would have. Haiku earns that back by the sixth request. Alternate between them and neither cache ever warms.

A Mycroft report checks the cache before it suggests any model change.

Cumulative input cost over ten requests: staying on Sonnet 4.6 against switching to Haiku 4.5 Staying costs $0.045 a request and reaches $0.45 after ten. Switching costs $0.177 on the first request, then $0.015 a request, reaching $0.31 after ten. The lines cross between requests five and six. $0 $0.20 $0.40 123 456 789 10 requests after the switch pays back on request 6 request 1 on Haiku: 3.9× the cost of staying stay on Sonnet 4.6 switch to Haiku 4.5
$0.045
a request on Sonnet 4.6, cache already warm
$0.177
the first request on Haiku 4.5, paying to write its cache
$0.015
each request after that, once Haiku's cache is warm
$0.426
a request if you alternate, so neither cache is ever warm
Input cost of one agent step at list prices: about 141,000 cached prefix tokens and 1,000 fresh ones, five-minute cache.

Each finding shows its arithmetic, and what it can't see.

Finding 2 from the sample report, with a reviewer's notes.

FINDING 2  21.0% of spend is the agent's own
  housekeeping, on claude-opus-5

  What your metrics backend shows
  - No cache on claude-opus-5 for this traffic,
    so moving it to claude-haiku-4-5 costs
    nothing to switch and saves $3,096.98
    over this period.

  At most $3,096.98 over this period
    It is a ceiling, not a saving - see below.

  What your metrics backend cannot tell you
  - Whether your agent lets you choose a model
    for this traffic separately from the main
    one.
  1. Priced by kind of request.

    Claude Code labels each request as someone's work, a subagent, or its own background calls. Mycroft prices each kind separately.

  2. The cache is checked first.

    This traffic has no cache to lose, so the switch costs nothing. Where it does, the finding says how many requests it takes to earn back.

  3. A ceiling, never a promise.

    Every dollar figure is an upper limit on your own totals at published prices.

  4. What the data can't show is printed.

    Each finding ends with the question only your team can answer, instead of a guess.

One read-only credential, then a report.

  1. You create one credential that can only read metrics.

    You make it, you scope it, you can revoke it. Mycroft never gets write access to anything.

    Grafana Cloud
    Access policy token, scope metrics:read, on one stack
    Datadog
    Application key, scopes metrics_read and timeseries_query

  2. Mycroft checks what your telemetry can answer.

    Before it reads any usage, it lists the labels it found and what each missing one costs you. A gap becomes a stated limit, not a quiet error.

      [OK] Token counts        claude_code.token.usage
  [OK] Input/output split  type
  [OK] Model names         model
  [OK] Cache split         cacheRead, cacheCreation
  [OK] Developer identity  user.email
  [--] Repository
       Spend cannot be attached to a codebase,
       so the question 'is this where our work
       is meant to be?' goes unanswered.
  [OK] Request segments    query_source, effort
  [OK] Reported cost       claude_code.cost.usage

  3. We send the report. A month later, what moved.

    The next report compares two equal windows: findings that appeared, findings that stopped, and how each ceiling moved. It never calls a smaller month a saving.

    SINCE LAST TIME
  Comparing 2026-08-29 to 2026-09-25 against
  2026-08-01 to 2026-08-28.
  $16,920.34 against $18,412.60 - $1,492.26
  less, 8%.

  No longer firing
  - 21.0% of spend is the agent's own
    housekeeping, on claude-opus-5

It reads daily totals. Nothing it could change.

19
models priced, from Anthropic, OpenAI and Google
31 days
after the last price check, Mycroft refuses to print a report
1
read-only credential, which you can revoke at any time
0
prompts, responses, files or logs read

Reads

  • Daily token counts by model and type: input, output, cache reads, cache writes
  • The developer and request labels Claude Code attaches to them
  • Your backend's own cost figure, to check the arithmetic against

Never touches

  • Prompts, responses or code
  • Logs, traces or files
  • Anything it could change

Five design partners get a free report on their last two weeks.

Mycroft is early. If your team runs Claude Code and sends its telemetry to Grafana or Datadog, or could, write in and we will set it up with you.

Ask for a report

or write to hello@mycroftcompute.com