Describe what you want. Go home.
Sandcastle ships it.
Which provider. Which model. Which agent. What happens when it fails. What it costs. Where the data lives. Who approves what. You used to build all of that yourself. Now you describe what you want and go home. European-built, open source.
$ git clone https://github.com/gizmax/Sandcastle.git && cd Sandcastle && uv sync && uv run sandcastle serve
You became a developer to build things.
Not to manage API keys. Not to debug rate limits at 3 AM. Not to explain to your EU client why their data just passed through a Virginia data center. Somewhere between the framework docs and the third provider outage this month, you forgot what you were actually trying to build. Sandcastle remembers.
- ✕Three months building glue code before the first real feature
- ✕Locked into one provider. Pricing changes? Start over.
- ✕Invoice arrives. Surprise. No one tracked the costs.
- ✕Provider goes down at 3 AM. Your phone rings.
- ✕EU client asks where their data lives. You check Slack.
- ✕Compliance audit next month. Nobody documented anything.
- ✓Describe it, ship it. Tell AI what you need - in production today, not next quarter.
- ✓7 providers, one setting. Switch from Claude to Mistral in one line. Run locally with oMLX.
- ✓You know what it costs. Per provider. Per workflow. Per step. Right now.
- ✓Provider down? You sleep. Auto-failover switches to the next one.
- ✓EU data stays in EU. One toggle. Hard enforcement. Not a policy doc.
- ✓Agents as building blocks. 15 templates. Or describe what you need. AI designs the agent.
- ✓Audit trail from day one. Tamper-evident. EU AI Act ready. Always on.
Everything you need for production agents
From sandbox execution to production-grade orchestration - everything you need in one package.
DAG Workflow Engine
Define multi-step pipelines in YAML. Dependencies, parallel branches, data passing between steps.
Parallel Execution
Steps at the same DAG layer run concurrently. Fan out over lists with configurable concurrency.
Sandshore Runtime
Purpose-built agent runtime with circuit breaker, pool management, health caching, and production-grade optimization.
Pluggable Sandboxes
Four backends: E2B cloud microVMs, Docker containers, Cloudflare Workers edge, or local subprocess. Switch with one env var.
20 Step Types
llm, http, code, condition, classify, loop, race, sensor, gate, transform, notify, delegate, parse, openclaw. Mix AI with $0 deterministic steps.
Universal Advisor
Any LLM powers all AI features - Claude, Mistral, OpenAI, Ollama, Google, MiniMax. One config, short aliases like sonnet, opus, openai/gpt-4o, ollama/llama3.
Smart Auto-Failover
Automatic fallback on 429 or 5xx. Per-key cooldown tracking, ordered failover chains. SLO-aware routing picks the best model for critical ops, cheapest for simple ones.
AI Workflow Generator
Describe what you need in plain English. sandcastle generate creates a complete workflow YAML with steps, dependencies, and model selection.
Agent Memory - No OpenAI Required
Persistent context across runs powered by Anthropic Claude + local fastembed/ONNX embeddings. Semantic search auto-injects relevant memories into prompts. Zero OpenAI dependency.
Cost-Latency Optimizer
SLO-based model routing with feedback loop, model degradation alerts, and automatic severity-based recommendations.
AutoPilot v2
Thompson Sampling variant selection, Welch's t-test significance testing, and progressive rollout (canary 10% -> partial 50% -> full 100%).
Cost Intelligence
Per-provider cost breakdown with proactive savings recommendations. Shows "Would cost $45 via Mistral" comparisons so you always pick the right model for the budget.
Policy Engine
Declarative rules for PII redaction, secret blocking, cost guards. Applied per step or globally.
Budget Guardrails
Cost tracking per step and per run. Set hard budgets and get alerts before they blow up your invoice.
HITL Approvals
Pause workflows for human review. Approve, reject, or edit the data before the next step runs. Multi-strategy gates.
Circuit Breaker
Automatic failure detection with CLOSED/OPEN/HALF_OPEN states. Prevents cascading failures across your pipeline.
Progressive Rollout
Deploy experiment winners safely: canary (10%) to partial (50%) to full (100%). Statistical significance required before each stage advance.
EU Data Residency
Toggle one setting - all AI processing stays in EU or local. Built-in compliance with data residency requirements, not just a promise. Works with any provider that offers EU endpoints.
63 Tool Connectors
Slack, GitHub, OpenAI, Anthropic, AWS S3, Google Sheets, Airtable, Discord, Supabase, Pinecone, Stripe, Shopify, PagerDuty, Datadog, Langfuse, Qdrant, GCS, Azure Blob, Exa, and 44 more. Add tools: [slack] to any step.
A2A Protocol
Google's Agent-to-Agent protocol. Agent card discovery at /.well-known/agent.json, JSON-RPC 2.0 task management.
AG-UI Streaming
CopilotKit's Agent-User Interaction protocol. Real-time SSE streaming of agent state, tool calls, and text deltas to any frontend.
MCP Server
Built-in Model Context Protocol server. Run workflows, check status, and manage schedules from Claude Desktop, Cursor, or Windsurf.
Webhooks + SSE
Real-time event streaming via SSE. Webhook dispatcher for external integrations. Live updates, no polling.
Named Connections
Multiple credential instances per tool. slack:engineering, postgresql:analytics - named connections resolve to the right credentials automatically.
Browser RPA
Five modes: Playwright (selector-based), Computer Use (vision AI), DOM Extract (accessibility tree), LightPanda (10x faster headless via CDP), Browserbase (cloud-hosted, zero cold-start). Action caching, CAPTCHA escalation, execution replay.
AI Connectors
OpenAI, Anthropic, ElevenLabs, Tavily, Firecrawl, Pinecone. Chain multiple AI providers, vector search, web scraping, and text-to-speech in your workflows.
DevOps Connectors
Vercel, Cloudflare Workers, Datadog, PagerDuty, AWS S3, Redis. Deploy, monitor, alert, and manage infrastructure from workflow steps.
Full CLI Suite
templates, runs, replay, fork, approve, reject, generate, doctor. Global --json flag for scripting.
Real-time Dashboard
Runs, costs, schedules, dead letters, approvals, experiments, policy violations - all in one place. Visual workflow builder included.
236 Templates
118 built-in + 118 community templates. Marketing, sales, engineering, support, HR, legal, and general AI. Full Community Hub with one-click install and uninstall.
Evaluation Framework
A/B test models and prompts per step. Automatic quality evaluation and best-variant deployment with AutoPilot.
REST API + Docs
Full OpenAPI spec. Interactive docs at /api/docs. Zero-config local mode - sandcastle init + sandcastle serve.
Community Hub
Browse, install, and share workflow templates. One-click install from the dashboard or sandcastle hub install author/name from CLI.
OpenTelemetry
Optional OTLP instrumentation with workflow and step-level spans. Includes cost, duration, and token counts as span attributes. Install with pip install sandcastle-ai[otel].
Workflows as YAML
No SDKs, no boilerplate. Declare your pipeline, Sandcastle handles the rest.
name: "Lead Enrichment"
default_model: sonnet
steps:
- id: "fetch"
type: http # $0 - no LLM needed
http_config:
url: "https://api.example.com/company/{input.id}"
method: GET
- id: "enrich"
type: llm # single API call, no sandbox
depends_on: ["fetch"]
prompt: "Research: {steps.fetch.output}"
- id: "route"
type: classify # LLM-based routing
depends_on: ["enrich"]
classify_config:
categories: [hot, warm, cold]
input: "{steps.enrich.output}"
branches:
hot: [priority-outreach]
warm: [nurture-sequence]
cold: [archive]Mix agent + lightweight steps. One file.
Not every step needs a full AI agent. Use http for API calls ($0), code for Python snippets,
condition and classify for branching, loop for iteration, race for parallel competition,
sensor for polling, transform for templates, notify for alerts - and llm or full agent steps where you actually need AI.
- ✓21 step types - standard, llm, http, code, condition, classify, loop, race, sensor, gate, transform, notify, delegate, approval, sub_workflow, parse, openclaw, map, retry, switch, agent
- ✓Smart branching -
condition,classify, andraceroute to different paths - ✓$0 steps - http, code, condition, transform, notify cost nothing - no LLM call
- ✓Data passing - reference prior outputs with
{steps.id.output} - ✓Per-step models - Claude, OpenAI, MiniMax, Gemini - mix per step
- ✓Automatic retries - exponential backoff on failure
- ✓Persistent storage - results saved to disk (local) or S3 (production)
See everything. Control everything.
Runs, costs, schedules, dead letters, approvals, experiments, policy violations - all in one place. Includes a visual workflow builder for drag-and-drop pipeline design.
Talk to Sandcastle from your AI editor
Built-in MCP (Model Context Protocol) server. Claude Desktop, Cursor, and Windsurf can run workflows, check status, and manage schedules - all from the chat interface.
{
"mcpServers": {
"sandcastle": {
"command": "sandcastle",
"args": ["mcp"]
}
}
}8 tools. 3 resources. Zero config.
Add one JSON block to your client config. The MCP server connects to a running
sandcastle serve instance and exposes the full workflow API.
- ✓run_workflow - run a saved workflow by name with optional input
- ✓run_workflow_yaml - run a workflow from inline YAML definition
- ✓get_run_status - detailed status with all step results
- ✓list_runs - browse runs with status and workflow filters
- ✓cancel_run - stop a queued or running workflow
- ✓save_workflow - save YAML workflow definitions to the server
- ✓create_schedule / delete_schedule - manage cron schedules
- ✓Resources - read-only access to workflows, schedules, and health
236 workflow templates, ready to deploy
Skip the blank page. Pick a template, tweak the prompts, and ship. Browse the Community Hub or install via CLI.
Your workflows improve themselves
Set a goal. Walk away. Come back to a faster, cheaper, more accurate workflow.
Enterprise-Ready Compliance
Built-in EU AI Act compliance, tamper-evident audit trail, and PII redaction - so you can deploy AI in regulated industries.
Ready to ship agent workflows?
Install Sandcastle in under 30 seconds. No infrastructure, no setup - just works.
Boring tech, reliable results
No exotic dependencies. Battle-tested tools you already know. Local mode needs zero infrastructure.
Free forever. Seriously.
Sandcastle is fully open source under the Business Source License 1.1. Free for non-production use, converts to Apache 2.0 in 2030. No feature gates, no "contact sales" buttons. Every feature you see on this page ships in the free version - because there is only one version.
Buy me a coffee
I built Sandcastle in my free time and I plan to keep improving it - new features, better docs, bug fixes, community support. If it saves you time or you just think it's cool, a coffee goes a long way.
It keeps me caffeinated and motivated to ship the next update.
☕ Buy Me a CoffeeCheck out my other projects
Apps I built and shipped. Give them a try.
Cardigo Wallet
Your loyalty cards, digitized. Scan barcodes, snap photos, store everything locally. Works offline, no account needed.
Try Cardigo →EYEON
Scientifically-designed vision training exercises. Neuroplasticity-based workouts for eye health and post-surgery recovery.
Try EYEON →Hush
Sleep timer for Mac. Set it and let your computer dream of electric sheep. Auto-sleep after movies or podcasts.
Try Hush →