USE CASE 01 / FIRST SAFE WRITE
ActionDock puts deterministic policy, owner approval, credential brokering, and an execution receipt between any existing AI agent and one real API write.
THE API-WRITE BOUNDARY
ActionDock is a supervised API-write gateway, not an open proxy and not an agent runtime. The destination, credential, and owner-authored transaction rules remain controlled outside the calling agent's context. This boundary applies only to writes routed through ActionDock; a direct login, credential, native connector, or browser path bypasses it.
The owner enters the HTTPS base URL, authentication type, methods, and a write-policy mode. The default uses path prefixes only; the recommended mode combines those prefixes with an owner-authored reasonableness rule. The constrained setup assistant can draft manual fields from typed or dictated instructions, but a draft never saves itself or authors the runtime rule.
Dashboard → API connectionsThe agent calls integration.execute with a connection, method, path, query, and body. ActionDock creates a job before any external request is sent.
integration.executeActionDock enforces the destination, method, public-network and SSRF protections, and execution limits on the server. Path prefixes also apply unless the owner explicitly selected advanced Reasonableness only mode. The connection version, resolved target, exact query, and body are bound into the transaction preview.
policy → allowed | rejectedA signed-in owner reviews and approves the exact transaction preview. In an AI-enabled mode, ActionDock then applies the owner's reasonableness policy as a probabilistic allow-or-reject check. It fails closed, does not search the web, and never receives the stored credential. Only an allowed request is sent with the server-held credential.
awaiting_approval → queued → runningActionDock records status, timestamps, safe response headers, bounded body, and errors. If the provider may have accepted the request but the response was lost, the job becomes execution_unknown. Blind retry stays blocked; an operator must reconcile downstream state before deciding what happens next.
job.succeeded | execution_unknownFIRST SAFE WRITE PATTERNS
Move one approved work item to its next state without exposing the system credential.
Write reviewed metadata only to permitted resource paths and preserve the provider response.
Show the correction payload, require an owner decision, and stop for reconciliation when execution is uncertain.
MODEL-INDEPENDENT / OWNER-APPROVED