> ## Documentation Index
>
> Fetch the complete documentation index at: https://lithi.ai/llms.txt
>
> Use this file to discover all available pages before exploring further.

# Start with Compute

## Start with one small result

Begin with a result you can judge from a small input. For example, ask your agent to classify eight
fictional product reviews as positive, negative, invalid or abstained, keep each supplied ID and
explain the evidence. Your acceptance rule can require all eight IDs, six accepted labels, one
invalid item and one abstention.

Keep the input small enough to review completely and use data you are allowed to process. Lithi owns
internal partitioning and Mac placement; you define the outcome, semantic constraints and policy.
Use [Choose work](/docs/compute/choose-work) if you are still deciding whether the task belongs on
Lithi.

## Check fit before access

Use [Usage fit](/docs/compute/usage-fit) to decide whether offloading the work helps now. Use [Cost fit](/docs/compute/cost-fit) when you have a comparable current baseline and need a cost comparison.

A fit assessment is not account access, a quote, approval or execution.

If you want a planning comparison before setup, use [Get a free estimate](/compute#assessment). It uses the
work you authorize to assess fit; copying the prompt does not install anything, upload data, request
a key or execute work.

## Confirm customer access and client

Use [Account and access](/docs/compute/access) to confirm the account, role, key/scopes and limits for the work.

Then [set up Compute](/docs/compute/setup-prompts) or [choose a compatible supported client](/docs/compute/clients). If you already use Claude Code, Codex, Cursor or another supported frontier agent, you can use the [customer AI setup guide](/docs/compute/clients/frontier-agent-setup).

Customer AI access is separate from provider Mac authority in Lithi Connect.

## Prepare the work and checks

Choose the [workflow](/docs/compute/workflows), freeze the [input](/docs/compute/inputs) and decide what verification you need.

Prepare the instructions with the input. Ask your agent to record the goal, the evidence and
dependencies it must keep together, the rubric and missing-context rule, the expected result and
the acceptance check it will use after the result returns. [Build a workflow request](/docs/compute/workflows/build)
explains this task brief. The Kit can prepare the permitted data and supported request; your agent
keeps the wider context and reviews the result.

If the work is time-sensitive or should run later, review [Booking and priority](/docs/compute/booking) before requesting the final quote.

## Fund and run the first test

When the task is prepared, buy the initial $25 credit yourself through the secure human payment
step. Your agent keeps the selected account and prepared request while you pay, then checks confirmed
funding before requesting the final quote and approval. Later top-ups are voluntary. The shared
[credit and payment guide](/docs/compute/billing/invoices-and-payment) explains payment recovery and
the separate purchase invoice and usage statement.

When your account, client, workflow, input, verification rule and timing are ready, continue with
[First test](/docs/compute/first-test). That guide covers the signed quote, required approval, one
submission, result review and reconciliation. Scale only after the small result passes your checks
and every exception has a clear owner.