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

# Use Lithi with Claude Code

Claude Code can help you assess a bounded task for Lithi Compute and, after you connect an approved
account, use the customer Work tools within your permissions. Your agent keeps its context and you
review the result and decide what to do with it.

## Assess the work first

Ask Claude Code to identify a bounded part of your task, its expected output and the checks you
would use to judge it. For example, provide a small set of recorded diagnostics you are authorized
to share and ask for recurring failure patterns with references to the supplied records. Compare that
with keeping the review in Claude Code; [choose work](/docs/compute/choose-work) explains what fits.

You can start with this brief:

> Assess whether these supplied diagnostics fit a supported Compute operation. Propose a small test
> that identifies recurring failures and cites each supporting record. Show the input, output and
> checks first; ask before transferring data or spending.

Moving suitable work can preserve some subscription allowance for other tasks. If you pay directly
for API usage or purchased credits, compare the equivalent costs instead; a subscription allowance
and direct charges are different constraints. Claude Code's reasoning, review and tool output still
consume host usage. Check [Anthropic's current cost guidance](https://code.claude.com/docs/en/costs)
for your plan and usage.

## Connect the intended account

Start with [Compute access approval](/docs/compute/availability). Once your account is approved,
follow [Claude Code's MCP documentation](https://code.claude.com/docs/en/mcp) for the host's
current setup. The project configuration location is `.mcp.json`; its existence does not authorize
an edit or an account connection. Use the account-issued endpoint and a protected credential reference
supported by the host. Do not put a secret in a prompt, example, command argument or project file.

Confirm the authenticated account and scopes are the ones you intend before proceeding. A skill or
configuration file provides guidance or settings; neither grants access, permission to transfer data
or approval to spend. The [MCP guide](/docs/compute/clients/mcp) explains the connection and scopes.

## Check the tools and try a small test

Verify the host exposes exactly five customer Work tools: `lithi.quote`, `lithi.run`, `lithi.batch`,
`lithi.status` and `lithi.result`. Quote prices a bounded request; run and batch request execution;
status and result read a job's state and result metadata. Reporting and account administration are
separate surfaces.

Discover a supported operation and prepare the permitted diagnostics in its accepted format. If the
account needs credit, complete the secure browser funding step yourself and confirm the credit on the
same account. Request a quote for the bounded test, then obtain the required transfer and spend
approvals before submitting. Even a synthetic submission can use paid credit. [Your first test](/docs/compute/first-test)
walks through the decision.

## Inspect and recover

Use the job's existing ID with `lithi.status` and, when ready, `lithi.result`. Result metadata is not
the decrypted output: retrieve authorized files, verify their digest and check that each finding cites
a supplied diagnostic and describes a recorded failure. See [result verification](/docs/compute/results/files-and-verification).
If a submission or read times out, inspect the same job and retain the original intent and idempotency
key; do not create a new job merely to recover an uncertain outcome. Follow the returned next action.