Claude, models and budgets

What you’ll learn

• The ways your organisation can reach Claude, and which one to pick — including prepaid Livra Claude credit.
• How Claude does the work.
• How models are chosen per phase.
• How the daily, per-run and turn limits protect your budget.

Where it lives

Settings › Claude and cost belongs to your organisation alone: its own Claude account, its models and its budget. Owners and admins change it; members can look. No other organisation sees or uses your keys.

1 · How Livra reaches Claude

Under How Livra reaches Claude, pick one:

• API key — your organisation's Anthropic API key. The simplest choice.
• Amazon Bedrock, Google Vertex AI, Microsoft Foundry — your company's cloud. Non-secret values such as the region go in Provider settings, one per line.
• Company LLM gateway — its address in Provider settings, its token in the secret field.
• Demo (no Claude calls) — for trying the app; nothing real is built.

Keys and tokens are stored encrypted. Afterwards the page only shows that one is saved and its last 4 characters.

Tip: Claude subscriptions (Pro, Max, Team, Enterprise) can't be used for background runs. Use an API key or your company's cloud account.

Or: Livra Claude, prepaid

When your platform offers it, Whose Claude runs your work at the top of the page lets you skip the account entirely: choose Livra Claude, prepaid. Whoever manages billing buys a pack of credit — with PayPal, mobile money or a card — and each run takes its real Claude cost from it. A pack costs its credit plus 1% for payment fees, and comes with an invoice. You're told when the credit runs low; when it's used up, runs stop until you top up. Your models and limits below still apply, and credit left over is kept if you switch back to your own account.

2 · How Claude does the work

Livra runs Claude in the background: steps run by themselves, and people follow along and approve pushes in the page. It is billed per use on the account above.

3 · Models

Under Models, choose a model for each group of phases — Set up, Build and chat, Plan, and Check and Document — and a fallback for when one is busy. A common pattern: the most capable model for Plan, where a good decision saves the most work later, and a faster one for Check and Document. Anyone can still pick another model for their own runs in the Claude panel.

4 · Budgets

Example: Acme Logistics has ten developers and wants to stay under 40 USD a day.

1. Most per day (USD): 40. Only your organisation's runs count towards it.
2. Most per run (USD): 4, so one run that goes wrong can't eat the day.
3. Most turns per run: keep the default unless runs stop too early.
4. Click Save.

The cards at the top show today's spending, the provider, the per-run limit and whether sessions are reused. Recent runs with cost shows what each run cost — look there first when the day's budget runs out early.

5 · Savings

• Reuse the Claude session of a feature lets later phases continue the same conversation, so Claude doesn't reread everything. Keep it on.
• Let Claude search the web and Let Claude start sub-agents are off by default. They rarely help, and sub-agents multiply cost.

Running steps keep the settings they started with; new steps use the new ones.

Tip: change one setting at a time and watch Recent runs with cost for a day. You'll see the effect clearly.

Check yourself

1. An admin saves an API key. Can they read it again later?

No. It's stored encrypted; the page only shows that it's saved and its last 4 characters.

2. A run is in the middle of Build when you lower the per-run budget. Does it stop?

No. Running steps keep the settings they started with; the new limit applies to new steps.

3. Why are sub-agents off by default?

They rarely help and they multiply the cost of a run.

4. You use Livra Claude and buy a pack with 50 USD of Claude use. How much do you pay, and what happens when the 50 USD are used?

50.50 USD (the credit plus 1% for payment fees). When it's used up, runs stop until you buy another pack.