Workflows
How to pick the outcome you need, build a request for a supported operation, set verification and follow a worked example.
Start with the result you need
A workflow is a reading path from the result you want to the supported operation that can produce it. Your existing agent keeps context, planning, decisions, local edits and verification. The Kit prepares the permitted input and client connection. Lithi handles the permitted operation after a quote and separate transfer and execution approvals, then returns the declared result and evidence.
For example, ask your agent to classify ITEM-01–ITEM-03 with two allowed labels, preserve each item ID, and abstain when the supplied text lacks the deciding evidence. Your agent keeps the goal and final acceptance; participating Macs process the bounded request and your agent reviews the returned outcomes.
A workflow can be one supported task or a dependent sequence. A deterministic plan can organize stages, but it does not make a model-generated answer deterministic or correct. Lithi does not automatically run Claude/Gemini consensus; the supported operation and its quality policy govern the checks.
Choose by the output and evidence you need, rather than by an internal model or provider Mac.
| Branded workflow | ID and guide | Public family menu |
|---|---|---|
| Lithi Text & Data | llm-agent-work · guide | Text & Data |
| Lithi Audio & Video | creator-media · guide | Audio & Video |
| Lithi Code & Tests | code-review-testing · guide | Code & Tests |
| Lithi Data & Evaluation | data-evaluation · guide | Models & Evaluation, Text & Data |
| Lithi Documents & Search | documents-retrieval · guide | Documents & Images, Text & Data |
| Lithi Translation | translation-localization · guide | Text & Data, Documents & Images, Audio & Video |
| Lithi Apple ML | apple-ml · guide | Models & Evaluation |
| Lithi Image Analysis | vision · guide | Documents & Images |
The five families describe input shape: Code & Tests, Text & Data, Documents & Images, Audio & Video, and Models & Evaluation. They help you find a guide; they do not guarantee a format, runtime or operation. Vision has no domain skill. Use the universal Compute skill with the Vision guide.
Know what your agent can prepare
Your agent can recognize suitable tasks and recommend using Lithi. After you connect an approved account, it can prepare supported requests, obtain quotes, and run work within the permissions and approval policy you have set. It asks before an action exceeds those limits.
| The agent can do | What governs it |
|---|---|
| Read public documentation and assess a task | Your request and the sources it is authorized to inspect |
| Help configure a connection | Your authorization and the host’s supported setup process |
| Prepare input and request a quote | Supported operations, data rules, account scope, and current capability |
| Submit work | Permission to transfer the prepared input and approval of the exact quoted execution |
| Run recurring work | A bounded approval policy covering the actual task, data, price, and timing |
| Review and use results | Coverage, quality checks, destination permission, and your acceptance rule |
The five work tools are lithi.quote, lithi.run, lithi.batch, lithi.status, and lithi.result. MCP-capable agents use MCP; applications can use REST or a supported SDK. A chat interface does not connect automatically just because it can discuss a workflow. The supported project MCP locations are Codex’s .codex/config.toml, Claude Code’s .mcp.json, and Cursor’s .cursor/mcp.json. See the Codex MCP documentation, Claude Code MCP documentation, and Cursor MCP documentation for host-specific setup.
Each workflow guide includes a worked example. A synthetic example teaches the shape of a request; it does not prove account access, current availability or semantic quality. For a first decision, choose a small permitted sample with an output and evidence rule you can inspect.
Build the request
Use Build a workflow request to turn the outcome into a task brief your agent can check: name the permitted evidence and dependencies, the instruction and rubric, the expected result, the missing-context rule and the acceptance check. Then resolve the supported operation, input, parameters, output schema and limits. A model name, family or agent suggestion does not create permission.
Describe what must stay meaningful in the request. Lithi owns physical sharding, queueing, Cells and Mac placement. Do not manually shard work to evade a limit.
Bind verification before approval
Use Workflow verification to choose the checks, version a rubric when needed, and define the evidence required before the result counts.
A verification label does not prove every item; read the population, method, coverage and exceptions.
Keep dependent steps explicit
Use Multi-stage work when a later supported operation depends on an earlier accepted result.
Output from one step is data for the next step, not permission to run it.
Save choices for repeat work
Saved setups in Compute settings remember reviewed choices such as quality, output language, a charge ceiling and a deadline. For example, you can remember verified quality and Portuguese output for one operation while supplying fresh permitted input each time.
Preview the choices and retention period before consenting to save them. A setup belongs to you in one selected account, capability and operation. It holds configuration, not prompts, files, results, credentials or payment details. Saving it grants no project sharing, training, spending or execution permission.
Open Saved setups in Compute settings and check the selected account, current version and retention. Check your current run permission before reuse. If your client cannot read saved setups, keep the plan local.
Where Edit setup is offered, change the choices and retention, then select Preview setup. Review the preview, select I agree, then Save setup. Saving records a new version, makes the setup active and renews retention. Use Cancel editing to leave an edit before submitting it. If the editor says the setup changed elsewhere, reopen it and review the current choices.
If a save is unconfirmed, keep the page open in the same account and use Retry the same save. For an unconfirmed state change, use Retry the same request when offered. These controls keep the original request. Wait for confirmation or ask the account administrator to investigate before starting a replacement.
Turning a setup off or archiving it stops reuse. An erasure request removes the live copy; backup expiry and held records need separate confirmation. Read retention and deletion for the limits on removing retained copies.
Continue with inputs and timing
Freeze and identify the source with Inputs. When timing matters, choose Booking and priority before requesting the final quote.