OPTIONAL · REAL CLAUDE · SUPERVISED

Connect a real model.
Keep every safety gate.

A practical nine-step guide from a free dry run to one paid API request, independently checked evidence, explicit human approval and a local draft export.

Your API key stays localDrafts onlyNo auto-retries

Follow this on your own computer. AI Teachers does not collect API keys, provide a paid model endpoint, or grant external tool access.

Illustrative person planning a project beside a laptop — Illustrative image, not a screenshot of a paid AI call.

BEFORE ANY CHARGE

Real API usage is optional and may incur fees.

Set your own spend limits in Anthropic Console, use permitted inputs, and understand that local budget reservations are estimates. Even an independent quote check cannot prove every generated statement is true.

Download full Python kit Try free browser lab Open illustrated workbook

STEP 01

Get your practice files

Download the starter ZIP and unzip into a local folder you control. Keep source fixtures fictional. The real Claude runner is separate from the free simulator.

python3 -m unittest test_agentic_os.py test_agentic_os_claude.py

STEP 02

Run the no-cost browser and CLI simulations

The browser cockpit requires no account or model key. The original Python simulator also makes zero network requests.

python3 agentic_os.py --state ./demo-state run --task demo-safe.json

STEP 03

Choose and budget the model

Start with a model that performs well on your verified task. Use Sonnet 5.5 by default. Fable 5.1 is more expensive and may be refused by the conservative teaching policy.

python3 agentic_os_claude.py --state ./live-state plan --task demo-safe.json --model claude-sonnet-5-5

STEP 04

Configure API access privately

Create a key in your own Anthropic Console, then set the ANTHROPIC_API_KEY environment variable in your terminal or protected secrets manager. NEVER paste it into AI Teachers, a prompt or a webpage.

export ANTHROPIC_API_KEY="your-private-key-set-locally"

STEP 05

Run one paid, draft-only request

After reviewing the admission plan, opt in explicitly. The API runner sends source notes only to Anthropic and requests structured JSON with NO external tools.

python3 agentic_os_claude.py --state ./live-state run --task demo-safe.json --model claude-sonnet-5-5 --confirm-paid-api-use

STEP 06

Inspect verified evidence

Open the local ledger. It checks claimed excerpts against original source text, validates their IDs and records token usage. A checked quote is not proof the summary is accurate.

python3 agentic_os_claude.py --state ./live-state report

STEP 07

Record a real human decision

Inspect the draft and source facts yourself. The script never auto-approves. The reviewer label is a learning convenience, NOT a secure production identity.

python3 agentic_os_claude.py --state ./live-state review --task-id feedback-safe-001 --reviewer learner --decision approve

STEP 08

Export a local draft, not a public action

The approved artifact can be exported to a NEW Markdown file. No sending, scheduling, posting, publishing or repository writes occur.

python3 agentic_os_claude.py --state ./live-state export --task-id feedback-safe-001 --output approved-feedback-draft.md

STEP 09

Reconcile costs before scaling

Review actual provider charges, test contradictory and hostile notes, and complete the 30-day rollout workbook. Do not rely on the local estimate for a global account bill.