What each client includes
Which AI tools and developer interfaces you can use with Compute, and what each one gives your agent.
Choose the client you already use
Compute gives your existing agent a harness for supported work. Your agent keeps its model, goal, project context, edits and final decisions. The Compute Kit provides guidance and local modules that help prepare permitted work and inspect results; authenticated MCP or a supported SDK or REST integration exposes the current operations.
Your agent can recommend a suitable task and prepare a supported request. The client sends that request to Compute, while the account, operation, data rules and approval policy govern what happens. Writing a prompt alone does not create a new supported operation.
Choose the language client for Compute API calls, then check which local helpers it supplies. The language package and the supplied Compute Kit are related but different surfaces: a Python client does not automatically include the TypeScript privacy module, and an SDK helper is not an additional MCP work tool. For request-building decisions, see Build a Compute workflow request.
Skills teach instructions and examples. They do not add model weights, credentials or permissions.
| Client | Guide |
|---|---|
| Claude Code | Use Lithi with Claude Code |
| Codex | Use Lithi with Codex |
| Cursor | Use Lithi with Cursor |
| Another compatible agent | Use Compute with your AI |
| MCP client | MCP |
| Direct HTTP integration | REST |
| Python | Python |
| TypeScript | TypeScript |
| Swift | Swift |
| Go | Go |
| .NET | .NET |
| Java/Kotlin | JVM |
Use this list as the compatibility check. Choose only a client and protocol combination that the current release documents as supported.
“Any agent” means a compatible host and supported transport with approved access. It does not mean that every agent has the same tools, local Kit modules, model, or account permissions.
Inspect the connection before changing it
The Setup module can discover a registered host, produce an exact connection plan and run no-spend
diagnostics over the configuration your caller supplies. Its planHostInstall, inspectSetup,
doctor and renderInspectReport helpers show what the host supports and what a change would do.
Read that result before applying a setup change. A local CLI, a cloud agent workspace and a managed
deployment can expose different configuration files and scopes, so a helper in one SDK family does
not imply parity in another.
Setup and diagnostics do not approve paid work, transfer input or submit a job. After the connection is understood, the agent still chooses the task, checks the quote and approval, and reviews the returned result.
What each client includes
The matrix records the wire and local coverage documented for each released language client. Use it before choosing an SDK so you know which preparation, workflow and result boundaries remain in your application.
| Capability | TypeScript | Python | Go / Swift / .NET | JVM |
|---|---|---|---|---|
| Compute wire client | Yes | Yes | Yes | REST guidance |
| Complete four-module Kit | Yes | No | No | No JVM SDK |
| Record and format helpers | Data module | Selected adapters | Caller-supplied | Caller-supplied |
| Pseudonymization and restoration | Data module | Caller-supplied | Caller-supplied | Caller-supplied |
| Compile, plan and resume workflows | Core + Workflows | Caller-supplied | Caller-supplied | Caller-supplied |
| Result artifact handling | Full Kit delivery path | Selected bundle reader and local-destination helpers | Caller-supplied | Caller-supplied |
The four TypeScript Kit modules are @lithi/kit-core, @lithi/kit-data,
@lithi/kit-workflows, and @lithi/kit-setup. Each package exposes its public API from .;
the supplied package names do not imply public package-registry installation. Use the exact local
release and each package's root entry point; consult that release's reference for the delivery
entry point.
The Python package includes bounded record and format adapters, selected result-bundle readers,
local-directory destination checks, and schema-backed getImpactReport. It does not provide the
complete TypeScript preparation, workflow, or delivery steps. Go, Swift and .NET are wire
clients; bring local preparation, workflow coordination and result delivery from your application
or another reviewed library. JVM guidance describes direct REST integration; it is not a JVM SDK
package.
For the component reading order, see Kit contents and Use the Kit with your agent. Use the Kit helper reference for exact exports. For a concrete support-ticket preparation flow, read What leaves your machine, then review result verification and result delivery.
Keep customer and provider authority separate
A customer agent can quote, submit and read customer Compute work within its scopes.
A provider agent manages authorized Macs through Lithi Connect. The same frontier agent can help a person in both roles when it holds each role's separate authority, but those credentials and permissions stay separate.
Keep reporting separate from work tools
The customer work surface has five MCP tools: lithi.quote, lithi.run, lithi.batch,
lithi.status and lithi.result. A monthly impact report uses separate permission and an approved
reporting API or file path; it is not a sixth work tool. Read your monthly Compute impact report
and use the Impact Reports API when your account grants
that reporting access.
Connect before paid work
A setup prompt or client configuration doesn't approve paid execution. Verify the account and transport first, then request a quote for the actual work and obtain the required approval.
Use Set up Compute for the guided setup boundary, Reference for exact API contracts, and Kit contents for the supplied local modules and workflow guides.