> ## 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: Set up your own AI agent
description: >-
  Give your own AI agent the Compute tools, using the shared assessment and delegation
  prompt rather than one you write yourself. Approval stays with a person.
route: /docs/compute/clients/frontier-agent-setup
page_id: docs-compute-clients-frontier-agent-setup
page_type: howTo
content_layer: guided documentation
surface_profile: guided_docs
audience: People delegating Compute work to a frontier coding or chat agent
voice: D
reader_question: How do I let my own AI agent assess, quote and run Compute work safely?
primary_action: Confirm the setup with one small assessment
source_locale: en-US
source_status: APPROVED
source_version: docs-compute-clients-frontier-agent-setup-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"
---

## What you need first

You need an agent that can load MCP tools, and a work key for it to use. Codex, Claude Code and Cursor all qualify, as does any other MCP-capable client.

A work key holds six work scopes and nothing else. Management scopes are never mintable on it, so an agent set up this way cannot invite members, change destinations, or touch finance.

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

## Use the shared assessment and delegation prompt

There is one canonical prompt for assessment and delegation. Copy it from the portal rather than writing your own.

It teaches the order that matters: assess whether the task fits, request a quote, bring you the price, and submit only after you approve. A prompt written from scratch usually drops the approval step, because nothing in the tool list forces it.

Paste it into your agent's own instruction or system prompt field. Keep it as the current version, and re-copy it when your agent starts behaving as though a quote were permission.

## Add the tools without guessing a config path

Each agent keeps its tool configuration in its own place, and those locations change between releases.

Find the location in your own tool's documentation. Do not copy a path from this page, from an older article, or from another team's setup.

Point that client at Compute's MCP surface, then confirm the tool list is exactly `lithi.quote`, `lithi.run`, `lithi.batch`, `lithi.status` and `lithi.result`. A sixth work tool did not come from Lithi.

## Keep approval with a person

Your agent can assess, quote and read results all day without spending anything. A quote carries an expiry and a signature, and it is explicitly non-authorizing.

Approval is a separate act, and it belongs to you. Give the agent the quote to bring you, never the decision to make.

Nothing about the agent's instructions changes this. Spend limits, scopes and approvals are enforced where work is admitted, not inside your agent.

## Confirm the setup with one small assessment

Ask your agent to assess one small, well-bounded task and bring you a quote. Do not approve it yet.

A correct setup returns a fit assessment and a signed, expiring quote, and nothing has run. That is the whole success condition.

If your agent returns a completed result you never approved, stop and revoke the key. Then check which prompt and which tool list the client actually loaded.

Read [how to connect an MCP client](/docs/compute/clients/mcp) for transport and authentication checks, and [why approval is always separate from a quote](/docs/compute/concepts/approval-and-authority).