Features

Features

Workflow engine

DAG execution

Every workflow is a directed acyclic graph. Steps declare dependencies via depends_on and the engine automatically parallelizes independent branches. No manual orchestration needed.

15 step types

From simple LLM calls to parallel races and human approvals. See Step Types for the full reference.

Natural language input

Describe a workflow in plain language (any language - Czech, English, Japanese, anything) and the generator builds a complete YAML definition with steps, dependencies, and model assignments.

YAML precision

When you need full control, write YAML directly. Define models per step, set timeouts, configure retries, declare input schemas with validation.

Variable interpolation

Reference inputs with {input.field_name} and previous step outputs with {steps.step_id.output}. The engine handles resolution automatically.

Input schemas

Define typed input fields with descriptions, defaults, and required flags. The dashboard renders forms automatically from the schema.

---

Execution environment

4 sandbox backends

Backend	Use case	Isolation
E2B	Production cloud	Full VM-level isolation
Docker	Self-hosted	Container with seccomp + CapDrop ALL
Cloudflare	Edge execution	V8 isolate
Local	Development	Subprocess (no isolation)

Multi-model support

7 models across 4 providers with automatic failover:

• Claude (Sonnet, Opus, Haiku)
• OpenAI (Codex, Codex Mini)
• Google (Gemini 2.5 Pro via OpenRouter)
• MiniMax (M2.5)

Provider failover: automatic retry on 429/5xx with per-key cooldown and ordered fallback chains.

56 connectors

Slack, GitHub, Salesforce, Stripe, SAP, PostgreSQL, MongoDB, browser automation, and 48 more. See Integrations.

---

Dashboard (20 pages)

Overview

Real-time metrics: active runs, success rate, cost tracking, recent activity feed with live SSE updates.

Workflow builder

Visual DAG editor with drag-and-drop step creation, dependency wiring, model selection per step, and live YAML preview. Template browser for starting from existing workflows.

Template gallery

118 built-in templates organized by packs (Marketing, Sales, Engineering, etc.), searchable, with preview and one-click install to workflows directory.

Run monitoring

Real-time step-by-step progress with SSE streaming. View inputs, outputs, token usage, cost per step. Compare runs side-by-side. Export to PDF.

Scheduling

Cron-based scheduling with visual cron builder. Pause, resume, delete schedules. View next run time.

Approvals

Human-in-the-loop approval queue. Approve, reject, or skip pending approval steps. Configurable timeout with auto-approve/reject fallback.

Integrations management

Browse all 56 tools by category. Configure credentials per tool. Test connections. View which workflows use each integration.

API keys

Create, list, rotate, and delete API keys. Set expiry dates. Configure IP allowlists per key. Grace period for key rotation.

Settings

Runtime configuration, license status, theme selection. Persistent settings saved to database.

System health

Backend status, database connectivity, queue health, sandbox availability, memory system status.

Cost optimizer

Track spending across models and workflows. Automatic model downgrade suggestions. Budget alerts.

AutoPilot

A/B testing for workflows. Deploy experiments, track metrics, auto-advance rollouts based on performance.

Evaluations

Run eval suites against workflows. Compare quality across model changes. Track evaluation scores over time.

Dead letter queue

Failed workflow runs that can be retried, inspected, or resolved. Automatic DLQ routing for transient failures.

Policy violations

Guardrail violations log. See which steps triggered policy rules, with severity and details.

---

CLI (27 commands)

Full command reference at CLI Reference. Highlights:

sandcastle init              # Interactive setup wizard
sandcastle serve             # Start API + dashboard
sandcastle run workflow.yaml # Run a workflow
sandcastle generate "..."    # Generate from natural language
sandcastle doctor            # Diagnostics (7 sections)
sandcastle hub search        # Browse community templates

Global --json flag for scripting and CI/CD integration.

---

Agent memory

Persistent memory across workflow runs:

• Write admission control - Importance scoring prevents trivial memories
• Novelty detection - Deduplication via word overlap heuristics
• Decay / TTL - Memories expire based on age and relevance
• Graph support - Optional Neo4j for relationship-aware memory
• Structured enrichment - Auto-extracted keywords and tags
• API + CLI - Full CRUD for memory management

---

Protocols

Protocol	Standard	Use case
A2A	Google Agent-to-Agent	Agent discovery via /.well-known/agent.json, inter-agent task delegation
AG-UI	CopilotKit	SSE streaming for real-time UI updates in compatible frontends
MCP	Anthropic Model Context Protocol	Expose Sandcastle as a tool server for Claude and other MCP clients

---

Enterprise features

Security

• Fernet credential encryption (AES-128-CBC + HMAC-SHA256)
• API key rotation with configurable grace period
• IP allowlisting per key (IPv4 + IPv6)
• Docker seccomp profiles with CapDrop ALL
• Distributed rate limiting (Redis sorted sets)
• Security headers (CSP, X-Frame-Options, HSTS-ready)
• SSRF prevention, path traversal protection
• CORS wildcard filtering

License system

• Ed25519 offline signature verification
• Three tiers: Community, Pro, Enterprise
• No features blocked - purely a compliance/legal signal
• License info in dashboard, CLI doctor, and API

Workflow versioning

• Git-like version management (draft, production, archived)
• Promote/rollback versions via API or dashboard
• Diff between versions

Webhooks

• Configurable webhook delivery on run events
• Retry with exponential backoff
• Callback URL support per run

PDF reports

• Export any run as a formatted PDF
• Step-by-step results with metadata
• 1,257 lines of PDF generation logic