Pre-built Colab notebooks for every FiGuard pattern. Each runs against the live sandbox — no signup, no local setup. Click Open in Colab, then Runtime → Run all.

Budget types

Notebook	What it demonstrates	Open
monetary-single-agent.ipynb	Single agent with a $500 spend cap — authorize → confirm → deny lifecycle plus velocity controls and shadow mode
monetary-with-categories.ipynb	Per-category allocations ($300 flights / $200 hotels) — spending in the wrong category is denied
monetary-fleet.ipynb	Multi-agent fleet with delegation tokens — each sub-agent gets a hard cap; one runaway doesn't affect the others
agency-client-cost-report.ipynb	Per-client cost attribution for an agency — category caps, a caught loop, and a client report built straight from the ledger; optional webhook forward to your reporting/observability stack
tokens-single-agent.ipynb	Token budget that caps LLM token usage instead of dollars — same authorize/confirm/deny lifecycle
tokens-with-categories.ipynb	Token budget with per-model limits — GPT-4o and Claude get separate token pools
api-calls-single-agent.ipynb	API call budget that caps external data-source requests — useful when each call costs money

Framework integrations

Notebook	What it demonstrates	Open
openai-agents-travel-booking.ipynb	OpenAI Agents SDK — @guarded_function_tool wraps a flight-booking tool; shows authorized and denied calls
langchain-shopping-agent.ipynb	LangChain — FiGuardCallbackHandler in the callback stack gates every tool call before it executes
crewai-research-fleet.ipynb	CrewAI — FiGuardCrewGuard wraps crew tools; Part 1 shows a single agent, Part 2 shows a fleet with delegation tokens

Agent incident playbook

Real-world failure modes and how FiGuard stops them.

Notebook	What it demonstrates	Open
agent-incidents/01_infinite_loop.ipynb	Agent stuck in a loop makes hundreds of calls — budget exhaustion stops it automatically
agent-incidents/02_duplicate_payment.ipynb	Retry logic fires the same payment twice — idempotency key makes the second request a no-op
agent-incidents/03_concurrent_overspend.ipynb	Two agents race to spend the same budget — FiGuard serializes authorization so only one wins
agent-incidents/04_rogue_subagent_fleet.ipynb	One sub-agent in a fleet goes rogue — delegation token caps it without disrupting the other agents
agent-incidents/05_category_violation.ipynb	Agent tries to spend outside its allowed category — category enforcement blocks the attempt

Advanced framework scenarios

Side-by-side before/after: the same agent workflow without and with FiGuard.

Notebook	What it demonstrates	Open
framework-scenarios/01_langchain_payment_retry.ipynb	LangChain payment retry loop — without FiGuard a bug retries indefinitely; with FiGuard the retry budget stops it
framework-scenarios/02_langgraph_research_loop.ipynb	LangGraph research loop — each graph iteration is pre-authorized; the loop stops the moment the budget is exhausted
framework-scenarios/03_langgraph_supervisor_fleet.ipynb	LangGraph supervisor with three sub-agents — delegation tokens give each agent a hard cap; the rogue researcher can't drain the fleet
framework-scenarios/04_crewai_parallel_crew.ipynb	CrewAI parallel crew — a runaway market researcher is capped at $2; the financial analyst and report writer complete unaffected

Sandbox

All notebooks connect to the shared demo environment — no account needed.

• Key: sb_live_demo (built into the SDK default)
• Dashboard: figuard-sandbox-g1ha.onrender.com/ui

Each notebook prints a direct link to its own budget in the dashboard after running, so you can inspect the spend tree without searching.

The sandbox resets periodically. No uptime SLA — for production use, self-host.