Compute services and limits

Start with the result

Start with the outcome you need. Use the account's capability summary and matching release material to check the exact operation, accepted input, expected output and limits. A service name alone does not establish what can execute.

For example, choose “extract invoice fields with source-page evidence” or “compare a declared set of model outputs against a rubric.” Then use the account's capability summary and matching release material to resolve the exact capability, operation, input and output. Do not convert a marketing label into an invented operation identifier.

For example, an available capability row tells you that discovery found a matching option for the account at that moment. It still does not quote a charge, admit a job, or prove that a result was delivered; resolve the operation and output contract before proceeding.

At the REST boundary, a job names one capability and one operation.

Choose an information family and workflow

Lithi's public papers and AI-agent guidance use one small work model:

Input → Service → Workflow → Execution → Verified output

• Input is the permitted information you want processed. For orientation, think in five broad families: Text & Data, Documents & Images, Audio & Video, Code & Tests, and Models & Evaluation. These are browsing labels, not format promises.
• Service is the exact capability + operation that performs one bounded kind of work. Human-friendly service behavior can be described as Extract, Analyze, Transform, Generate, Compare, or Verify, but the callable identifier still comes from discovery and the matching release material.
• Workflow connects one or more services to reach the requested outcome. Translation, for example, is a service/workflow behavior that can apply to several kinds of input; it is not an input family.
• Execution is Lithi's responsibility: planning suitable parallel work, selecting eligible execution, retrying safely, and checking what counts. Customers and AI agents do not choose Macs, cells, queues, physical shard sizes, or retry workers.
• Verified output is returned through the selected operation's declared output contract. The result may be text, structured data, vectors, media, reports, analysis of supplied build/test evidence, or another supported artifact only when that contract declares it.

The five MCP work tools are controls over this process, not services: lithi.quote, lithi.run, lithi.batch, lithi.status and lithi.result. Use discovery instead of relying on a static service list.

Do not describe Lithi as accepting “any format.” The exact accepted representations, limits, output contract, verification requirements, and current availability belong to the selected service release.

The five families and workflow labels help you choose a reading path; they do not promise an operation or availability. Confirm the exact service operation before preparing a large input.

Discover the capability

GET /v1/capabilities uses account access with jobs:read. Its OpenAPI operation ID is listCapabilities; the TypeScript method is listCapabilities(limit, cursor).

The response's data contains items, catalog_digest, and an optional next_cursor. Each item has:

Use limit between 1 and 100 and follow returned cursors unchanged. Discovery is read-only. An available row is not a quote, admission or evidence of a completed run.

Resolve the operation contract

Use the matching capability description supplied with your approved integration. Check public_operations, max_files, max_ciphertext_bytes, max_input_tokens, max_output_tokens, accepted data profiles, verification requirements, and billing unit.

The summary endpoint does not return a generic workflow parameters_schema or a complete output schema. Do not add guessed fields to compensate. If the matching description or output contract is missing, resolve it before quoting or sending data.

The operation name matters independently of the capability name. A verify operation checks the declared evidence; it must not be described as generating a transcript, rendered video, or new model merely because the capability label suggests that subject.

Keep request, transfer and execution limits separate:

All applicable constraints must hold. The largest transport allowance is not an execution allowance. Do not truncate required evidence or choose internal shard sizes to force a fit.

Read token prices with the metering recipe

Lithi measures text, code, documents and media in tokens. Each content type has a signed metering profile. One input rate and one output rate apply. Pages, images, minutes and native units are not separate customer prices. Item counts and recorded test or build counts describe workload and evidence, not billable token quantities. Native build, test, browser-test and release execution are outside customer Compute.

Review the signed quote's input/output tokens, maximum charge and policy before separate approval. Admission, execution, result acceptance and delivery remain distinct steps.

Recover a missing or changed capability

If discovery is unavailable, keep the request diagnostic and do not assume a cached list is current. If a release or manifest digest changed, resolve the new description before reusing earlier parameters.

If no listed operation can return the evidence your task needs, narrow or change the task rather than inventing a callable operation. Use workflow discovery to record a complete choice before quoting.