Invoke is live
The execution layer between AI agents and real tools is now generally available. Invoke intercepts every tool call your agent makes — Stripe, CRMs, Slack, Linear, internal APIs — and prevents duplicate charges, wrong-entity writes, and unknown-outcome retries before they land.New features
Execution runtime with four control primitives. Every tool call is classified, committed to a durable record, reconciled against live state when outcomes are unknown, and returned with a structured trace. Read How it works.Tool safety classes. Tools are labeledidempotent, write-once-checkable, irreversible, or approval-bound to control retry and gating behavior. See Tool safety classes.Reconciliation. When an attempt returns an unknown outcome, Invoke checks live source-of-truth state before allowing a retry and blocks duplicate side effects automatically. See Reconciliation.Approvals. Pause approval-bound actions for human review and revalidate state between approval and execution so stale approvals never fire. See Approvals.Execution records and traces. Every call writes a durable record before any side effect and returns structured proof of every decision, attempt, and policy gate. See Execution records and Traces.TypeScript SDK. Install @invoke/sdk, instantiate a client, and replace direct API calls with invoke.call(). No framework changes required. See the TypeScript SDK.CLI. Authenticate, scaffold projects, deploy to the runtime, and call tools from your terminal with @invoke/cli. See the CLI reference.REST API. Reconcile outcomes and verify state directly over HTTP. See /outcomes/reconcile and /state/verify.Guides
Walkthroughs for the most common integrations:- Payments — safe retries for Stripe and other write-once charges
- CRM updates — prevent wrong-entity writes
- Slack approvals — route approval-bound actions to reviewers
- Linear issues — idempotent issue creation and updates
- Webhooks — verify state before acting on webhook payloads