Changelog

Added

• Schema Registry for Bus topics: define JSON Schema contracts for message payloads. Validation runs automatically at publish time and rejects non-conforming messages.
• Schema management UI in the Bus section — create, version, and assign schemas to topics without leaving the platform.
• Schemas are optional per topic. Unschematised topics continue to accept any payload.

Added

• MCP Server execution modes — three options for hosted servers: Always On (Fargate task runs continuously), On Demand (starts on request, stops after 15 minutes idle), and Disabled. On Demand reduces cost for infrequently used servers.
• Live performance metrics in MCP Server detail view — request count, success rate, and average latency displayed in real time.
• Workflow Factory — describe a use case in plain language; the platform generates a complete workflow including steps, agent configuration, and connections.
• DataInspector panel in the workflow editor — click any data or action node to inspect its inputs and outputs in context without leaving the editor.
• Agent step prefill response — configure a structured output template to make LLM responses more consistent and easier to parse in downstream steps.

Added

• P1 Bus: a workspace-scoped event message bus with hierarchical topics (e.g. orders/created, payments/failed). Publish from workflows, agents, or external systems. Fan-out to multiple subscriptions automatically.
• Three delivery modes per subscription: Pull (SQS queue, caller polls), Workflow (trigger a workflow run with the message payload), and Webhook (outbound HTTP POST to any endpoint).
• Bus API keys — credentials scoped to specific topic patterns for publish and/or subscribe access. Separate from workspace API keys.
• Inbound webhooks — external systems can publish directly to the Bus via HMAC-authenticated HTTP endpoints without a JWT token.
• Bus MCP server — agents and Claude Desktop can publish to and subscribe from Bus topics via standard MCP tool calls at /bus/mcp.
• Bus Trigger node in the workflow editor — create and manage Bus subscriptions directly inside a workflow definition.
• JSONPath payload filters — subscriptions can filter messages using JSONPath expressions (e.g. $.amount > 100) so only matching messages trigger delivery.
• Dead-letter queue support and replay API for failed message deliveries.
• Bus metrics endpoint — track message volume, delivery rate, and failure rate per subscription.
• Circular delivery protection — the platform detects and blocks infinite message loops automatically.

Added

• Notifications panel — bell icon in the top navigation bar shows live workspace notifications including run failures, approval requests, and workflow deployments.
• Command palette — press ⌘K (Mac) or Ctrl+K (Windows/Linux) to search across workflows, agents, runs, and settings from anywhere in the platform.
• Per-run audit timeline — the Audit page now includes a dedicated view per run showing every tool call, agent decision, data access, and approval event in chronological order.
• Configurable audit digest email — receive a scheduled summary of audit events at daily, weekly, or monthly frequency. Configure recipients and risk-level filters in Settings.

Added

• MCP Gateway Data Loss Prevention (DLP): gateway policies allow admins to configure tool allowlists and denylists per MCP server, and to apply input and output redaction rules before data reaches an agent.
• Rate limiting per MCP server — configure maximum request rates in Gateway settings to prevent runaway agent behaviour.
• Model Integrations in Settings — register AI provider API keys (Anthropic, OpenAI, Google, Azure OpenAI) at the workspace level. Agents reference a provider registration by ID rather than storing keys directly on each agent.

Added

• Connector logos displayed in the Connections list for faster visual identification.
• "Used by" indicator on each connection — see at a glance which workflows and skills reference a given connection.
• JSON workflow importer — import a workflow definition from a JSON file directly into the editor.
• Workflow template gallery — a curated set of starting-point workflow templates covering common use cases.
• CI/CD pipeline — Go lint, TypeScript typecheck, and automated test runs on every commit.

Added

• Workflows — build and publish agentic workflows as directed acyclic graphs. Trigger types: webhook, schedule, manual, event, API, and queue. Step kinds: trigger, data, skill, agent, MCP, logic, approval, action, notify, and storage.
• Agents — configure AI model instances with system prompts, trust levels (low, medium, high), and persistent key-value memory. Supported providers: Anthropic Claude, OpenAI GPT, and Google Gemini.
• Skills — package Lambda functions as reusable agent tools with JSON Schema input validation and sandbox execution. Categories: data, transform, API, and integration.
• MCP Servers — connect external and hosted MCP servers via the Model Context Protocol. All MCP traffic proxied through the platform gateway.
• Connections — OAuth 2.0 and API key integrations for Slack, GitHub, Jira, ServiceNow, Zendesk, Snowflake, PagerDuty, Datadog, Snyk, Okta, LinkedIn, and X.
• Approvals — human-in-the-loop decision gates embedded in workflows. Configurable risk level, SLA in minutes, reviewer assignees, evidence display, and editable approval payload.
• Audit log — immutable event log with risk classification (low, medium, high, critical) and 7-year retention.
• Datastores — S3-backed workspace object storage for agent data persistence.
• Secrets management — workspace-level credential storage with audited reveal and rotation.
• Workspace settings — named environments (production, staging, development), RBAC with three roles (admin, editor, viewer), workspace API keys with 16 granular scopes, and per-event notification preferences.
• Approval SLA monitoring — automated SLA breach detection and notification.
• Run debugger — step-by-step run inspection showing inputs, outputs, tool calls, logs, and errors per step.