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

---
title: Connect a client to Compute
description: >-
  Two entrances to Compute: delegate work to your own AI agent, or build directly
  against the API. Neither requires installing or enrolling hardware. Choose the
  client that matches how you work.
route: /docs/compute/clients
page_id: docs-compute-clients
page_type: hubIndex
content_layer: guided documentation
surface_profile: guided_docs
audience: Developers, integrators and agent operators choosing a Compute client
voice: D
reader_question: Which client should I use to submit Compute work?
primary_action: Set up your own AI agent
source_locale: en-US
source_status: APPROVED
source_version: docs-compute-clients-en-us-2026-09-05-v1
last_updated: '2026-09-05'
robots: index
claim_registry_pinned: true
claim_registry_resync: "npm run governed:tsx -- scripts/build-compute-public-claim-registry.ts --write"
---

## Two ways to reach Compute

Most people never write a Compute integration. They give their own AI agent the Compute tools and describe the work in plain language. The agent assesses fit, asks for a quote, and brings you the price to approve.

The second entrance is direct. You call the API yourself, over REST or through a released language client.

Neither entrance requires installing or enrolling any hardware. Helping power approved network tasks is a separate product, with a separate setup.

## What every client shares

The contract does not change when you switch clients. Four rules hold everywhere.

- The same seven request objects describe a task.
- A quote never authorizes work, and approval is always a separate act.
- Every response carries a typed state and a concrete next action.
- A result bundle is complete only when its manifest is present.

Your base URL and current API version appear in the portal under developer settings, and in the API description you download.

## Pick your entrance

| If you want to | Start with |
| --- | --- |
| Describe work in plain language and approve a price | Your own AI agent |
| Script or schedule work from a terminal | The command line |
| Build the flow into your own service | REST or a language client |
| Confirm what your client version can do | The compatibility matrix |

## Pages in this group

- **Your own AI agent** — give a frontier agent the Compute tools, using the shared delegation prompt.
- **MCP** — connect an MCP client and use exactly five work tools.
- **Command line** — the released commands, split by local act and remote effect.
- **REST** — one complete path from prepared input to verified result.
- **TypeScript** — typed request objects, bounded responses and result verification.
- **Python** — async waiting, cancellation and streamed bundle reading.
- **Go** — contexts, deadlines, typed errors and safe streaming.
- **.NET** — framework requirements, cancellation tokens and streamed results.
- **Swift** — package and platform requirements, and how this differs from the provider application.
- **Java and Kotlin** — what a released JVM surface must pass, and what to use today.
- **Compatibility** — client and protocol version, tested features, gaps and retirement dates.