AI orchestration that decides before the model does.
Switch models, keep everything.
AI orchestration is the coordination of models, tools and data so that one request becomes a finished result instead of a chain of copy-and-paste. Vodou is an AI orchestration layer that runs on your own machine. Before any model is called, Vodou matches the request against intent mappings for every skill and for the tools of every MCP server you connect; only when nothing matches does a model plan. It then executes the plan as a dependency graph — independent steps run together, joins tolerate a failed branch — across any model provider or local runtime, with your memory attached. Deterministic first, agentic second.
- Vodou is an AI orchestration layer that runs on your machine on macOS, Windows and Linux; a free Vodou account licenses the engine and your memory is stored on your computer.
- Vodou routes before inference: intent mappings for every skill and every MCP server tool are checked first, and a model is called only when no deterministic route matches.
- Vodou orchestrates every major LLM provider plus Ollama, LM Studio and a bundled llama.cpp behind one request, so the model can change without the skills, tools or memory changing.
- Vodou derives parallelism from the dependency graph of a plan instead of asking you to declare it; on one measured workflow chain that made the run 4.6× faster.
What is orchestration in AI?
AI orchestration is the coordination of multiple models, tools, data sources and steps so they work together on one request — deciding which model or tool handles each part, in what order, and what happens when a step fails. IBM defines it as coordinating "AI models, systems and integrations". Vodou does it locally, deterministically first, then with a model.
The enterprise version of the term (IBM watsonx Orchestrate, UiPath, Pega) means a control plane for a company's agents. The developer version (LangGraph, LiteLLM, n8n) means a framework, a router or a workflow engine you assemble. A Reddit thread that ranks for the phrase complains that "orchestration now means everything", and its top answer draws the useful line: deterministic orchestration (the code decides) versus agentic orchestration (the model decides). Vodou uses both, in that order.
Vodou runs on your machine and needs a free Vodou account so the engine can be licensed to you; after that the runtime binds only to localhost, model calls go straight from your computer to the provider you chose, and your memory is stored on your disk; a local model keeps every call on the machine too. The orchestration layer is personal: it coordinates the models, tools and memory you use, not a fleet.
Why does AI orchestration matter?
Without an orchestration layer you do the coordinating: you pick the model, paste the context, call the calendar, then mail, then Slack, one after another, and start over in the next tool. Capacity, not effort, caps what you get done. Vodou coordinates instead: it routes the request to a skill, a tool on any MCP server or a model from any provider, runs independent steps together, and asks you only at approval points.
The model is guessing about routing
Asked to “book Thursday with Sam”, a bare model may or may not call the calendar tool. Vodou matches the intent deterministically first, so the tool call is not a coin flip.
Serial by default
Most agents run one tool, wait, run the next. Vodou compiles a plan into a graph and moves independent steps into a together: block automatically.
Locked to one model
An orchestration layer inside one vendor orchestrates one vendor. Vodou sits above every major provider and local runtime; the model is a setting.
How does Vodou orchestrate models and tools?
Vodou routes each request deterministically before inference, lets a model plan only when routing finds nothing, compiles the plan into a dependency graph, and executes independent branches concurrently with approvals and memory attached.
- STEP 1
Match before you infer.
The request is scored against intent mappings for every skill and tool with one threshold table (exact, word-boundary and substring tiers, a routing floor, a confidence gap). A matched skill beats a tool; a tool beats a bare completion. The Console shows the route receipt so you can see why.
- STEP 2
Plan when nothing matches.
A model from your chosen provider — Anthropic, OpenAI, Google, Groq, DeepSeek, xAI, Mistral, Kimi, OpenRouter, Fireworks, Together, or Ollama, LM Studio and llama.cpp locally — writes the plan. You see it before it runs.
- STEP 3
Compile to a graph, run together.
Steps that read nothing from the step above move into a together: block; the plan card tells you it did. Tools across every MCP server you connect execute concurrently over pooled connections; a join records “2 of 3 settled” and continues when the count is met, naming the branch that died.
- STEP 4
Approve, remember, repeat.
Anything that sends, spends or deletes stops for approval on the channel you are in — Slack, Telegram, Discord or the Console. The result is distilled into local memory, so the next route starts with more context. Schedule it and it runs the same way unattended.
Deep dive: Automation
What can you do with Vodou's AI orchestration?
An LLM orchestrator for any model provider
One request, any hosted provider — Claude, GPT, Gemini, Grok, DeepSeek and more — or a local runtime; the model is a setting, and skills, tools and memory do not change when it does.
An LLM router that runs locally
Routing, provider keys and the memory the router consults all live on your machine; BYOK keys are never metered.
Deterministic routing before inference
Intent mappings for every skill and tool; a route receipt on every turn; edit a skill's trigger phrases and the route changes without retraining anything.
Parallel execution derived from the graph
No “run in parallel” flag to set; the compiler finds independence, and the join tolerates a dead branch.
Plan first, run on your say-so
Describe a multi-tool job and Vodou compiles a plan you can read before anything runs; approvals reach Slack, Telegram and every other channel you use as the same numbered prompt.
Open-source orchestration you can read
The gateway, the bundled MCP servers and every skill are Apache-2.0 on GitHub; the routing engine is proprietary and included with every plan.
Instead of moving through everything one task at a time, Vodou helps me advance multiple projects in parallel. I'm getting more leverage from the same amount of time.
Sources IBM: What is AI Orchestration? · IBM: What is LLM Orchestration? · Model Context Protocol specification · Vodou docs: workflows.md · Vodou docs: mcp-host.md · opper.ai: The 10 best LLM gateways in 2026
How does Vodou compare to LiteLLM, OpenRouter, n8n, LangGraph and watsonx Orchestrate?
LiteLLM and OpenRouter route model calls — LiteLLM self-hosted across 100+ LLMs, OpenRouter hosted across hundreds — and do it at a scale Vodou does not attempt. n8n and LangGraph orchestrate workflows you build; watsonx Orchestrate manages an enterprise's agents. Vodou orchestrates one person's request end to end: deterministic routing to skills and tools first, any model provider second, parallel execution and memory throughout, on your machine. If you need a team gateway or a visual builder, use them.
| CAPABILITY | VODOU public alpha | LITELLM model gateway | OPENROUTER hosted router | N8N workflow automation | LANGGRAPH agent library | WATSONX ORCHESTRATE IBM |
|---|---|---|---|---|---|---|
| What it orchestrates | A person's request: skills, tools, models, memory | Model calls (proxy/SDK) | Model calls (hosted) | Workflows + AI Agent nodes | Agent control flow (library) | Enterprise agents, one control plane |
| Routes to tools/skills before a model runs | Yes — intent mappings for every skill and tool | No | No | Your workflow decides | Your graph decides | No — the agent's LLM picks tools and collaborator agents |
| Models / providers | Any provider + Ollama, LM Studio, llama.cpp | “100+ LLMs” | “500+ models”, “80+ providers” | Chat-model nodes per provider | Any LangChain model | 14 providers via AI Gateway (OpenAI, Anthropic, Ollama…) |
| Fallbacks / load balancing across providers | Basic (provider is a setting) | Yes — retries, fallbacks, load balancing | Yes — automatic fallbacks | Fallback model on the AI Agent node | Node retry policies; model fallback via LangChain middleware | Not stated |
| Parallel execution | Derived from the dependency graph; tolerant joins | N/A | N/A | No — branches run one after another by default | Yes — parallel nodes per superstep, Send fan-out | Parallel nodes in agentic workflows (ADK) |
| Memory attached to every run | Yes, local SQLite with provenance | No | No | Memory sub-nodes per agent (Simple, Postgres, Redis…) | “Built-in memory stores conversation histories” | Agent memory across chat sessions |
| MCP | Client to unlimited servers + host | “Central MCP endpoint” | Hosted MCP server for its own data (models, prices, docs) | Yes — MCP client and server nodes | Yes — langchain[mcp] adapter (pre-release) | “A2A and MCP” |
| Approvals | Buttons in Slack/Discord/Console | No approval step; guardrails (PII, content filters) | — | Human review on AI tool calls (Slack, Telegram, chat) | “human-in-the-loop checks” | Human-in-the-loop in agentic workflows |
| Per-key / per-team budgets | Per-task budgets on the board | Yes — “per-key/team/user budgets” | Yes — spend limits per key and per member | — | — | Token-usage visibility; no budget cap stated |
| Runs where | Your machine, localhost only | Self-hosted | Hosted only | Self-host or cloud | Your Python/JS process | IBM Cloud, AWS, on-prem |
| Open source | Open-core; gateway, skills, servers Apache-2.0 | MIT, except an enterprise/ directory under a commercial license | Closed | Sustainable Use (fair-code) | MIT | Closed |
| Built for | One person | Platform teams | Developers, one bill | Ops/automation teams | Developers | “leaders responsible for AI at scale” |
| Price | $10/mo BYOK; 60-day trial | Free software (enterprise license paid); you pay providers | No markup on inference; 5.5% fee on card credit purchases | Free self-host; cloud from €20/mo | Free library | 30-day trial; Essentials from $530/mo, Standard from $6,360/mo |
- LiteLLM — docs.litellm.ai/docs/ · docs.litellm.ai/docs/proxy/users · github.com/BerriAI/litellm · github.com/BerriAI/litellm/blob/main/LICENSE
- OpenRouter — openrouter.ai/ · openrouter.ai/pricing · openrouter.ai/docs/faq · openrouter.ai/docs/guides/features/guardrails · openrouter.ai/docs/api-reference/api-keys/create-api-key · openrouter.ai/docs/guides/overview/mcp-server
- n8n — n8n.io/pricing/ · github.com/n8n-io/n8n · docs.n8n.io/flow-logic/execution-order/ · docs.n8n.io/advanced-ai/human-in-the-loop-tools/ · docs.n8n.io/integrations/builtin/cluster-nodes/sub-nodes/n8n-nodes-langchain.memorybufferwindow · docs.n8n.io/integrations/builtin/cluster-nodes/root-nodes/n8n-nodes-langchain.agent
- LangGraph — www.langchain.com/langgraph · docs.langchain.com/oss/python/langgraph/overview · docs.langchain.com/oss/python/langgraph/graph-api · docs.langchain.com/oss/python/langchain/mcp · reference.langchain.com/python/langgraph/types/RetryPolicy · docs.langchain.com/oss/python/langchain/middleware/built-in · github.com/langchain-ai/langgraph
- IBM watsonx Orchestrate — www.ibm.com/products/watsonx-orchestrate · www.ibm.com/products/watsonx-orchestrate/pricing · developer.watson-orchestrate.ibm.com/llm/managing_llm · developer.watson-orchestrate.ibm.com/tools/flows/overview · www.ibm.com/docs/en/watsonx/watson-orchestrate/base?topic=memory-using-agent-in-conversations
Frequently asked questions about AI orchestration
What is orchestration in AI?
Orchestration in AI is coordinating several models, tools and data sources so one request is handled end to end — which model or tool takes each step, in what order, and what happens on failure. Vodou is an orchestration layer for one person: it matches the request against intent mappings for every skill and tool first, calls a model from any provider only when needed, and runs independent steps together.
What is the best AI orchestration tool?
For an enterprise control plane, IBM watsonx Orchestrate; for a team's model gateway, LiteLLM; for visual workflows, n8n; for developers writing agent graphs, LangGraph. For orchestrating your own request across models, tools and memory on your machine, Vodou — deterministic routing, unlimited MCP servers, any model provider, parallel execution. Vodou is a public alpha with a 60-day free trial, so test it against your real tasks.
What is an LLM orchestrator?
An LLM orchestrator is the component that decides which language model handles a request, prepares its context, chains its calls and manages the result — the "backbone of the LLM app stack", in IBM's words. Vodou is an LLM orchestrator that runs locally: every major model provider plus Ollama, LM Studio and llama.cpp behind one request, memory injected as context, and a route receipt showing why each choice was made.
What is the best open-source AI orchestration platform?
For workflow engines, Conductor, Kestra and Flyte are Apache-2.0; for agent graphs, LangGraph is MIT; for model gateways, LiteLLM's core is MIT. Vodou is open-core: the gateway, the bundled MCP servers and skills are Apache-2.0 at github.com/VodouAI/OS, while the Rust routing engine is proprietary and included with every plan, including the 60-day free trial. If a fully open engine is the requirement, Vodou is not it.
What is an AI orchestration layer?
An AI orchestration layer is the software between the user (or app) and the models and tools, which decides where each request goes and assembles the answer. Vodou is that layer on your machine: it sits above any model provider and every MCP server you add, in front of the Vodou Console, Slack, Telegram, the terminal and an OpenAI-compatible API, and routes deterministically before a model ever runs.
What is the difference between AI orchestration and a harness?
Orchestration is the deciding — which model, tool or skill, in what order, in parallel or not. A harness is the whole runtime around a model: loop, tools, memory, sandbox and rules, of which orchestration is one part. Vodou is a harness whose orchestration is deterministic first: intent routing across every skill's and tool's mappings, then a model plan, then graph execution.
What is the difference between AI orchestration and agentic AI?
Agentic AI lets the model decide the next step; orchestration decides the steps from outside the model — code, rules, a graph. The Reddit consensus is "deterministic vs agentic", and the practical answer is both. Vodou routes deterministically when an intent matches, lets a model plan when it does not, and always executes the plan as a graph the model cannot rewrite mid-run.
Is MCP an orchestration layer?
No. MCP is a protocol that connects a model to tools and data; it says nothing about which tool to call, in what order, or in parallel. An orchestration layer uses MCP. Vodou is an MCP client to unlimited servers and an MCP host for Claude Desktop, Cursor and Zed, and adds the orchestration MCP leaves out: intent routing, a dependency graph, approvals, budgets and memory.
Orchestration is one part of a harness. Read more: AI harness · Agent harness.
More from Vodou
- AI memoryMemory you own that compounds with every session.
- AI harnessPlain English in, executed answer out.
- Agent harnessSkills, MCP, workers and hooks — already wired.
- Browser extensionYour chats, captured locally and typed back anywhere.
- SkillsWorkflows that run the same way every time.
- AutomationAgents that run unattended, within budgets you set.
- MCP gatewayOne MCP host for every editor on your machine.
- Local AI agentRuntime, memory and tools on your machine.
- Claude Code memoryMemory in Claude Code, Cursor and Codex on every prompt.
Decide before the model does.
Install Vodou, connect your models and tools, and watch the route receipt on your first request. Free for 60 days, no credit card.
Updated