A local AI agent that keeps its runtime, memory and tools on your machine.
Switch models, keep everything.
Vodou is a local AI agent: the runtime, the memory database and the tool layer all run on your computer and listen only on 127.0.0.1. It needs a free Vodou account so the engine can be licensed to you. Point it at a model running locally through Ollama, LM Studio or the llama.cpp runtime bundled with Vodou, and nothing you say leaves your machine — or bring your own API key for any model provider, and only the text you send that provider leaves; Vodou never meters those calls. It is free for 60 days, then $10 a month on your own keys. And it is more than a chat window: unlimited skills, MCP servers and messaging channels, plus a scheduler, all run against the same local memory.
- Vodou is a local AI agent whose runtime and memory live on your machine and listen only on 127.0.0.1; a free Vodou account licenses the engine, and with a local model through Ollama, LM Studio or the bundled llama.cpp nothing you say leaves your computer.
- Vodou runs fully local models through Ollama, LM Studio or a bundled llama.cpp, or your own API keys for any model provider — and Vodou never meters those keys.
- Vodou installs with chmod 600 on its memory database and .env, and every release passes a build gate that refuses to ship if the memory database contains a single row.
- Vodou is free for 60 days with no credit card, then $10 a month on your own keys; the gateway, skills, extension and MCP servers are Apache-2.0 on GitHub.
What is a local AI agent?
A local AI agent is an assistant whose runtime — the loop that plans, calls tools and stores memory — runs on hardware you control instead of a vendor's cloud. The model can be local too (Ollama, LM Studio, llama.cpp) or a cloud API you call with your own key. Vodou is local in the first sense and gives you both choices for the model.
"Local" hides two separate questions: where the model runs, and where your work lives. A chat app that calls a local model but stores your history in someone's cloud is local in the first sense only; an agent whose memory, tool credentials and schedules sit on your disk is local in the second, which is the one that matters when you switch vendors. Vodou keeps the second fixed and lets you change the first per task.
One thing to say before anything else about "local": Vodou requires a free account. The account gate is compiled into the engine, validates once and then keeps working offline, and exists so the engine can be licensed to you — it does not receive your memory, your conversations or your keys. Model calls go straight from your computer to whichever provider you chose, or never leave it at all with a local model.
Why does running an AI agent locally matter?
When the agent lives in a vendor's cloud, so does everything it learned about you — and it stays there when you switch to a better model, cancel the plan, or the vendor changes its memory policy. A local AI agent like Vodou keeps memory, credentials and schedules on your own disk, so the model is a swappable part: use Ollama for private work, a cloud key for hard problems, and the same memory for both.
Your context is the vendor's asset
Hosted memory belongs to one product; move to another and you start from zero. Vodou's memory is a SQLite file you can open, export or delete.
“Local” apps that phone home
A local model is not a local agent if the history, tool tokens and schedule are kept in a cloud. Vodou's gateway, MCP host and extension all bind to 127.0.0.1; there is no relay to turn on.
Local means weaker — unless you can mix
Small local models trail ChatGPT-class models on hard tasks. Vodou lets one request go to Ollama and the next to Claude or GPT on your own key, without changing anything else.
How does Vodou run as a local AI agent?
Vodou installs a Rust engine and a local gateway on your machine, connects to whichever model you choose, stores memory in a local SQLite database, and runs skills, tools, channels and schedules as local processes.
- STEP 1
Install on macOS, Windows or Linux.
The installer sets chmod 600 on .env, vodou-core.db and memory.db and makes the whole .vodou/ directory private to your user. Sign in with a free Vodou account; the 60-day trial starts, no credit card.
- STEP 2
Choose a model — local or your own key.
Vodou detects Ollama at localhost:11434 and LM Studio at localhost:1234, and ships its own llama.cpp runtime for machines with neither. Or paste a key for OpenAI, Anthropic, Google, DeepSeek, Mistral, Groq, xAI, OpenRouter, Together, Fireworks and more — plus any OpenAI-compatible endpoint — billed by the provider, never metered by Vodou. Switch per conversation.
- STEP 3
Let memory build locally.
Conversations from the Vodou Console, the Vodou Bridge extension on 22 AI sites, Claude Code and Cursor hooks, and every messaging channel you connect are captured into memory.db, indexed with FTS5 and vector search, and reranked on your CPU. Every fact keeps its source; you can correct, pin or undo it.
Deep dive: AI memory · Browser extension · Claude Code memory
- STEP 4
Act, locally.
Unlimited skills and MCP servers, a 60-second scheduler and a kanban board of agent workers all run as processes on your machine, reading the same memory. Anything that spends money or sends a message can require your approval first — in Slack, Telegram or wherever you are.
Deep dive: Skills · MCP gateway · Automation
What can you do with a local AI agent?
Run a local AI agent with Ollama
Pull a model in Ollama, pick it in Vodou, and chat, run skills and call tools with nothing leaving the machine. After the first sign-in the account check keeps working offline.
A local AI agent with memory that lasts
The memory is a SQLite database on your disk with full-text, vector and reranked search, not a context window that resets. Fix a fact once and every model you use gets the fix.
Local AI agent for Mac, Windows and Linux
Native builds for macOS 13+ on Apple Silicon and Intel, Windows 10/11 on x64, Linux on x64 and ARM64. The Linux build ships a systemd user unit.
A BYOK AI assistant that never meters your keys
Bring keys for any model provider; the provider bills you, Vodou charges a flat $10 a month for the platform after the trial. Keys are stored in a chmod 600 file on your machine.
A local AI agent harness, not just a chat
Skills with forced stopping points, any MCP server you add, an MCP host for your editors, a scheduler and a board of workers with budgets and approval gates. The person-facing view of that harness is at vodou.ai/ai-harness.
The same local agent in Slack, Telegram and every other channel
Message Vodou from WhatsApp, Discord, iMessage, Signal, Microsoft Teams or Google Chat — and install any other channel; the channel adapters run inside your local gateway, so the memory they read and write is the same local file.
Vodou helps me pick up where I left off without constantly re-explaining the context. I'm able to stay in motion and move my work forward much faster.
Sources Ollama documentation · llama.cpp on GitHub · Vodou AGENTS.md: Security defaults · Vodou licensing (NOTICE)
How does Vodou compare to OpenClaw, Hermes Agent and LocalAI?
OpenClaw, Hermes Agent and LocalAI are MIT-licensed, free forever and need no account; Vodou is free for 60 days, then $10 a month, needs a free account, and its Rust engine is proprietary. OpenClaw reaches 29 messaging channels, Hermes 27+; Vodou's are unlimited. What Vodou adds is memory that also works in your browser on 22 AI sites and in Claude Code and Cursor, an MCP host that shares that memory with your editors, and a per-client audit log.
| CAPABILITY | VODOU public alpha | OPENCLAW | HERMES AGENT |
|---|---|---|---|
| What it is | local agent + memory + harness | self-hosted personal assistant | self-hosted agent framework (Nous Research) |
| Free forever | no — 60-day trial, then $10/mo BYOK | yes — “no paid version” | yes (Nous Portal optional) |
| Needs an account | yes (free; licenses the engine) | no | no |
| License | gateway, skills, extension, servers Apache-2.0; engine proprietary | MIT | MIT |
| Local models | Ollama, LM Studio, bundled llama.cpp | Ollama and other local | Ollama, own endpoint |
| Cloud models on your key | Any provider, never metered | yes — hosted, subscription or gateway | Nous Portal, OpenRouter, OpenAI, others |
| Memory | SQLite + FTS5 + vector + reranker; provenance; correct / undo | “Persistent Memory” of preferences, projects, people | persistent memory, user profiles, FTS5 session search |
| Memory in the browser (ChatGPT, Claude…) | 22 sites via Vodou Bridge | no (browser automation, no chat-site extension) | no (browser tools only) |
| Memory in Claude Code / Cursor | hooks | no — openclaw mcp serve gives Claude Code its chats, not memory | no — runs inside VS Code, Zed, JetBrains via ACP |
| Messaging channels | Unlimited (installable) | 29 | 27+ (Telegram, Discord, Slack, WhatsApp, Signal, Teams, Matrix…) |
| MCP | Unlimited servers as client; host for your editors | client + mcp serve (chats and approvals) | client (stdio, SSE) |
| Audit of what a client read | yes (salted digests, 30 days) | no per-client log; security audit checks config | no per-client log; approval history in state.db |
| Binds to loopback only | yes, no setting to change it | loopback by default on a host install; configurable | configurable (VPS is the pitch) |
| OS | macOS, Windows, Linux | macOS, Windows, Linux | Linux, macOS, WSL2, Termux, Windows installer |
Sources checked (2026-09-17); competitor columns are as of that date.
- OpenClaw — openclaw.ai · docs.openclaw.ai · MCP · mcp serve · gateway security
- Hermes Agent — github.com/NousResearch/hermes-agent (MIT) · integrations · security · ACP
- LocalAI / LocalAGI — localai.io · github.com/mudler/LocalAI (MIT) · github.com/mudler/LocalAGI (MIT)
- Cline — cline.bot/pricing · github.com/cline/cline (Apache-2.0) · installing Cline · Memory Bank
Vodou column checked against the product on 2026-09-17.
Frequently asked questions about local AI agents
Is it possible to run AI agents locally?
Yes. A local AI agent runs its planning loop, tools and memory on your own computer; the model can be local too through Ollama, LM Studio or llama.cpp. Vodou runs this way on macOS, Windows and Linux, binds only to 127.0.0.1, and stores memory in a SQLite file on your disk. It needs a free account to license the engine; the account never sees your memory.
Is there a free local AI agent available?
Yes — OpenClaw, Hermes Agent and LocalAI are free and MIT-licensed. Vodou is not free forever: it is free for 60 days with no credit card, then $10 a month on your own model keys. What you pay for is the platform — memory that works in the browser, Claude Code and Cursor, unlimited skills, an MCP host and unlimited channels; the model itself stays local or billed by your provider.
Is local AI as good as ChatGPT?
For most hard reasoning tasks, no — a model that fits on a laptop trails GPT- or Claude-class models. The practical answer is to mix: run private or routine work on Ollama or LM Studio, send the hard problem to a cloud model on your own key, and keep one memory for both. Vodou switches per conversation across every major provider and the Ollama, LM Studio and llama.cpp runtimes without changing anything else.
Can you run a local AI agent with Ollama?
Yes. Vodou talks to Ollama at localhost:11434: pull a model in Ollama, select it in Vodou's model picker, and chat, run skills and call MCP tools with nothing leaving your machine. LM Studio works the same way at localhost:1234, and machines with neither can use the llama.cpp runtime bundled with Vodou.
What is the best local AI agent for Mac or Windows?
It depends on what you need it for. For a free chat-first assistant on many messaging apps, OpenClaw is the community favourite; for coding inside your editor, Cline. Vodou is the choice when you want one local memory across your browser, your IDE and your messaging apps plus skills and MCP tools, on macOS 13+, Windows 10/11 and Linux — free for 60 days, then $10 a month.
Do AI agents run locally?
Some do and most don't: ChatGPT, Claude and Gemini agents run in the vendor's cloud, and many “local” apps still store history or tokens remotely. Vodou's runtime, memory, MCP host and channel adapters all run as processes on your machine and listen only on 127.0.0.1. The only outbound calls are to the model provider you chose — or none at all with a local model.
What are the hardware requirements for a local AI agent?
The agent itself is light — Vodou runs on any macOS 13+ Apple Silicon or Intel Mac, Windows 10/11 x64, or Linux x64/ARM64 machine, and its memory search runs on the CPU. The model sets the real requirement: Ollama and LM Studio list RAM and GPU needs per model, and a cloud key on Vodou needs no local model at all.
What is a local AI agent harness?
A harness is the layer around a model that gives it memory, tools, skills and guardrails; a local harness runs that layer on your machine. Vodou is one: unlimited skills with forced stopping points, unlimited MCP servers, a scheduler, a board of workers with budgets and approval gates, and a local memory they all read. The person-facing view of that harness is at vodou.ai/ai-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.
- AI orchestrationRoute before inference. Run tools together.
- 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.
- Claude Code memoryMemory in Claude Code, Cursor and Codex on every prompt.
Switch models, keep everything.
Install Vodou on your Mac, Windows or Linux machine, point it at Ollama or your own keys, and let a local memory build from day one. Free for 60 days, no credit card.
Updated