Start with the job your agent needs to do.
Each guide gives you a small working workflow, the exact decision rule your agent should follow, and the limits you should know before production use.
Prevent a duplicate action
Give equivalent retries one stable identity, atomically reserve it, and execute only when the lock says yes.
Open quickstart →Wait for an outside event
Create a temporary webhook inbox, let the agent pause, then retrieve the event when it resumes.
Open quickstart →Require a human decision
Create a secure approval link and keep the agent blocked until the status is explicitly approved.
Open quickstart →Provide a veto window
Hold an explicitly permitted action, accept an exception-only veto, then atomically acquire one execution right.
Open quickstart →Check integration readiness
Validate supplied MCP metadata and block installation when required version, tools, protocol, or transport information is missing.
Open quickstart →Require N-of-M authorization
Create distinct member links, meet an independent threshold, and atomically consume one execution right.
Open quickstart →Keep a durable deadline
Derive due and escalation state from an external clock even when the originating agent is offline.
Open quickstart →Bound human-review capacity
Use per-subject caps, retry-safe reservations, and protected priority units for scarce review attention.
Open quickstart →Hand an incident to an operator
Create a private operator link and track acknowledgement, resolution, and deterministic SLA state.
Open quickstart →Pin observed terms before commitment
Record immutable terms, interface, version, and price evidence without storing the terms body.
Open quickstart →Preflight a provider policy
Require declared capabilities and reject forbidden ones with an exact, explainable match result.
Open quickstart →Bind an undo path before execution
Make a compensation fingerprint a prerequisite and track whether compensation remains due.
Open quickstart →Stop stale context reuse
Publish and check active, superseded, disputed, revoked, or expired digest status.
Open quickstart →Evaluate context use rights
Require an active rights envelope for the exact requested use before processing.
Open quickstart →Enforce a cumulative disclosure budget
Bound sharing across agents and sessions by total, recipient, and category counters.
Open quickstart →Check quality evidence before selection
Inspect and compare task-specific evidence for free, then contribute a privacy-safe outcome with a free verified key.
Open quickstart →Inspect a document before AI
Check a local file for hidden text, suspicious instructions, secrets, metadata, macros, and related signals.
Open workflow →Prepare untrusted text for an agent
Normalize, inspect, scan, detect secrets, optionally redact, and validate model output before use.
Open quickstart →Rewrite without another LLM call
Apply deterministic wording changes, verify protected facts, and use the result only when the preservation report passes.
Open quickstart →These are controls, not guarantees
ScrubMyText returns deterministic results and review signals. Your application still owns authorization, business rules, destination-system safeguards, and the final decision to act.