The AI harness that controls every model and every tool.
Never re-explain yourself to an AI again.
An AI harness is the software around a model that gives it a loop, hands, memory and rules — the part that turns "a chatbot answered" into "the work got done". Vodou is an AI harness that runs on your own machine and is built for a person, not a codebase. You type plain English into the Vodou Console, a Slack thread, ChatGPT or your terminal; Vodou routes it to the right model from any provider, the right skill from a library you can extend, or the right tool across every MCP server you connect, with your memory attached. The model can change tomorrow; the harness, and everything it has learned about you, stays.
- Vodou is an AI harness 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 each request before any model is called, across any model provider, unlimited skills and unlimited MCP servers, using intent mappings for every skill and tool.
- Vodou attaches the same persistent memory on every surface: 22 AI chat sites through the Vodou Bridge extension, Claude Code and Cursor, MCP clients such as Claude Desktop, Slack, Telegram, WhatsApp and any other messaging channel, the terminal and an OpenAI-compatible API.
- Vodou is open-core: the gateway, skills, extension and MCP servers are Apache-2.0; the Rust engine is proprietary.
What exactly is an AI harness?
An AI harness is the software around a language model that turns it into a worker: it runs the model in a loop, hands it tools, keeps its memory and enforces rules about what it may touch. Claude Code and Codex are harnesses around one vendor's model. Vodou is a harness around all of them, built for a person rather than a codebase.
Greg Isenberg put it plainly on X (13 Sep 2026):
"A harness does 4 things: 1. Runs the model in a loop so it keeps working step after step instead of answering once and stopping. 2. Gives it hands to read files, call tools, open portals, and run code. 3. Manages its memory so hour three of a job still knows what happened on hour one. 4. Enforces the rules about what it can touch and when it has to stop and ask a human."
His conclusion is the point of this page: "A wrapper was one model doing everything and a harness is a router. It gets better the more you use it." Wikipedia files the same idea under "agent harness, also known as agent scaffolding". (It is not Harness.io, the CI/CD company — same word, different product.)
Vodou needs a free Vodou account so the engine can be licensed to you; after that it runs on your machine, binds only to localhost, and keeps your memory in a database on your disk; the only text that leaves is what you send to the model you chose, and with a local model through Ollama, LM Studio or the bundled llama.cpp, nothing does. That is what makes it a personal harness: Claude Code is a harness for a repository, OpenClaw is a harness for a chat window, and Vodou is a harness for your whole working day — every model, every tool, every place you talk to an AI.
A shorthand both ChatGPT and Perplexity now use: an AI agent = a model + a harness. The model supplies language and reasoning; the harness supplies context, memory, tools, permissions, the execution loop and the audit trail. That also settles the framework question — an agent framework gives developers building blocks to create agents; a harness is the running environment that operates them, which is why Vodou is installed rather than imported.
Why does an AI harness matter?
Without a harness, you are the harness. You pick the model, paste the context, copy the answer into the next tool, and remember what you decided yesterday — for ChatGPT, then Claude, then Cursor, each with its own memory. Vodou takes that job: one place where every model provider, MCP server and skill you use is already connected to one memory, so the request is the only thing you type, and it stops resetting.
You are doing the routing
Which model is good at this, which tool has the data, which prompt worked last time. Vodou decides before inference, from what it knows about you.
Every AI is a separate silo
A decision made in ChatGPT does not exist in Claude or your terminal. A harness owns the memory; the models borrow it.
Tools run one at a time, by hand
Calendar, then mail, then Slack. A harness runs them together and asks you only when a human is needed.
How does Vodou work as an AI harness?
Vodou takes plain English from wherever you are, routes it deterministically before any model is called, executes with your memory and tools attached, and keeps what it learned for next time.
- STEP 1
Say it, anywhere.
Type into the Vodou Console, a Slack, Telegram or Discord thread, the vodou terminal client, or any of 22 AI chat sites through the Vodou Bridge Chrome extension. The same harness answers on every surface.
- STEP 2
Route before inference.
Vodou checks intent mappings for every skill and tool first — a skill wins over a tool, a tool over a bare model — so "book Thursday with Sam" reaches the calendar tool without a model guessing. When nothing matches, Vodou picks from any model provider you use or a local runtime (Ollama, LM Studio, bundled llama.cpp).
Deep dive: AI orchestration
- STEP 3
Execute with memory and tools attached.
Relevant facts from your local memory ride along; tools across every MCP server you connect run concurrently when independent; anything that sends, spends or deletes stops for your approval on whichever channel you are in.
- STEP 4
Remember, and improve.
The turn is distilled into facts in memory.db on your disk. Correct one and every model gets the fix. Repeated work can be drafted into new skills you review and promote — the harness gets better the more you use it.
Where does Vodou work?
Vodou's harness works everywhere you talk to an AI: in the browser on 22 AI chat sites through the Vodou Bridge Chrome extension, in Claude Code, Cursor and Codex through hooks, in Claude Desktop, VS Code, Windsurf, Zed and any MCP client through the MCP host, in Slack, Telegram, WhatsApp and every other messaging channel you connect, in the Vodou Console web chat, from the command line, and through an OpenAI-compatible API on your machine.
| SURFACE | HOW MEMORY GETS THERE | READS | WRITES | NOTES |
|---|---|---|---|---|
Browser — 22 AI chat sites ChatGPT, Claude, Gemini, AI Studio, Grok, Perplexity, DeepSeek, Copilot, Le Chat, Qwen, Kimi, Z.ai, T3 Chat, OpenRouter, Poe, Meta AI, Manus, You.com, Duck.ai, NotebookLM, HuggingChat, Character.AI | Vodou Bridge extension: Ctrl+B injects into the chat box; capture is automatic per site | ✓ | ✓ | Chrome Web Store · 35 host patterns · the panel prompts you to install Vodou if it isn't running · → Browser extension |
Browser — any text box on any page | Ctrl+Shift+B visible insert; “Fill this form from Vodou” (never submits) | ✓ | opt-in per site | Same extension · → Browser extension |
Claude Code | Four hooks: SessionStart, UserPromptSubmit, PostToolUse guard, SessionEnd extraction; plus a generated MEMORY.md in the workspace | ✓ | ✓ | → Claude Code memory |
Cursor | Same hook binary via .cursor/hooks.json, written by the installer | ✓ | ✓ | → Claude Code memory |
Codex in Cursor / VS Code, Agent Mode | Hooks + MCP | ✓ | ✓ | — |
Claude Desktop, VS Code, Windsurf, Zed, any MCP client | Vodou MCP host: vodou-core mcp install <client>; stdio or loopback HTTP 127.0.0.1:8787; per-client token + vault | ✓ (vault-scoped) | ✓ | → MCP gateway |
Any agent that speaks MCP | Vodou-Recall tools: search_memory, memory_store, memory_correct, memory_reject, memory_pin, memory_get, get_current_work | ✓ | ✓ | Plus feed tools that trigger automations |
Messaging — Slack, Telegram, Discord, WhatsApp, iMessage, Signal, Microsoft Teams, Google Chat, voice, web and more | Built-in channels plus any channel you install; same assistant, same memory; approvals reach every channel | ✓ | ✓ | WhatsApp: only threads you start, only your own messages |
Vodou Console localhost:8765 | Web chat with memory injected every turn; Memory tab; Brain map; correct / reject / pin / undo; Projects; Board | ✓ | ✓ | The inspect-and-fix surface |
Terminal — ./do, vodou-core mem | mem search, mem why, mem pin, mem correct, mem import, mem vault | ✓ | ✓ | A full, documented CLI |
Vodou CLI (TUI) | vodou from any directory: the same chat loop in your terminal, memory included | ✓ | ✓ | vodou -p “…” one-shot |
OpenAI-compatible API /v1/chat/completions on 127.0.0.1:8765 | Point any OpenAI-format client or app at Vodou and it answers with memory, skills and tools | ✓ | ✓ | Streaming supported |
Core HTTP API 127.0.0.1:8766 | /api/v2/memory/recall and the rest of a typed local API; TypeScript SDK | ✓ | ✓ | Bearer token, local only |
Kanban board workers | Each spawned agent gets the top memory matches in its prompt | ✓ | ✓ | → Automation |
Scheduler, automations, Skill Consoles | Scheduled and event-triggered runs execute with the same memory | ✓ | ✓ | → Automation |
Imports (one-way in) | ChatGPT, Claude, Obsidian, OpenClaw, Hermes, Letta, Vodou pack | — | ✓ | Facts keep an import:<source> label |
Local models and any hosted provider Ollama, LM Studio, bundled llama.cpp | Whichever model answers, it reads the same memory | ✓ | — | The model changes. Your memory doesn't. |
What can you do with an AI harness?
Switch models, keep everything
Every major model provider behind one request: Anthropic, OpenAI, Google, Groq, DeepSeek, xAI, Mistral, Kimi, OpenRouter, Fireworks, Together and more, plus any OpenAI-compatible endpoint. Change the model in a setting; memory, skills and tools stay.
An AI harness for local LLMs
Point Vodou at Ollama, LM Studio or the bundled llama.cpp and the same harness runs fully on your machine; BYOK keys are never metered.
Memory that follows you into every AI
Press Ctrl+B in ChatGPT, Claude or Gemini and Vodou types the relevant facts into the box, visible and editable before you send.
Skills that run the same way every time
Unlimited SKILL.md skills with inline steps the model cannot skip; install more from a sha256-pinned catalog or write your own.
Deep dive: Skills
Tools across every MCP server, run together
Calendar, mail, Slack, browser and files dispatched concurrently when independent; approvals reach the channel you are in.
Ask from wherever you are
Slack, Telegram, Discord, WhatsApp, iMessage, Signal, Teams, Google Chat, the terminal, or any app that speaks the OpenAI API.
Vodou lets me focus on the outcome instead of figuring out which model, tool, or workflow to use. It removes a lot of the manual coordination from AI-assisted work.
Sources Greg Isenberg on X, 13 Sep 2026 · Wikipedia: Agent harness · Google Cloud: What is an agent harness? · Model Context Protocol specification · Vodou docs: skills.md · Vodou docs: mcp-host.md · Vodou docs: memory-follows-you.md
How does Vodou compare to Claude Code, OpenClaw and OpenCode?
Claude Code, OpenClaw and OpenCode are real harnesses, each deeper than Vodou at its own job: Claude Code for a repository, OpenClaw for a personal assistant in 29 chat channels, OpenCode for terminal coding across 75+ providers. Vodou is the harness for the whole day — it sits across all three, adds memory that follows you into 22 AI chat sites, and routes to tools before a model runs. Vodou's trade-off: an alpha product with a required free account.
| CAPABILITY | VODOU public alpha | CLAUDE CODE Anthropic | OPENCLAW open source | OPENCODE open source |
|---|---|---|---|---|
| Built for | A person's whole AI day | A codebase | A personal chat assistant | Coding in the terminal |
| Models | Any model provider + Ollama, LM Studio, llama.cpp | Claude models — via Anthropic, or Bedrock, Google Cloud, Foundry on CLI/VS Code/JetBrains | “hosted and local model providers” | “75+ LLM providers”, including local models |
| Runs on your machine | Yes; localhost only; free account licenses the engine | CLI yes; web/cloud sessions optional; Claude subscription or Console account on most surfaces | Yes — “State, memory, and credentials live on your hardware” | Yes; “does not store any of your code or context data” |
| Memory | Local SQLite, hybrid search, provenance, correct/undo; follows you into 22 AI sites | CLAUDE.md + auto memory, per project | “Persistent Memory” on your hardware | AGENTS.md rules files; no built-in cross-session memory |
| Skills | Unlimited SKILL.md skills with steps the model cannot skip | Skills, plugins, hooks | Community skills, ClawHub, “can even write its own” | SKILL.md skills (also reads .claude/skills) |
| MCP | Client (unlimited servers) and host for Claude Desktop, Cursor, Zed… | Client; claude mcp serve exposes its own tools | Client registry; openclaw mcp serve exposes its conversations | Client (local + remote servers) |
| Messaging channels | Unlimited (installable) | Telegram, Discord, iMessage (Channels, research preview); Slack via @Claude | 29 (docs list 32 integrations) | — |
| Runs tools in parallel | Yes, from the dependency graph | Parallel sub-agents, background agents | Parallel sub-agents | Parallel subagents; multi-session |
| Approval before side effects | Yes, to any channel | Permission prompts; relayable to Telegram/Discord/iMessage (preview) | Exec approvals in Control UI, macOS app or chat | Allow / ask / deny rules |
| Open source | Open-core: gateway, skills, extension, MCP servers Apache-2.0; engine proprietary | Closed (“All rights reserved”) | MIT, all of it | MIT |
| Community size | Small (public alpha) | Very large | Very large (390K GitHub stars) | Very large (208K GitHub stars) |
| Price | $10/mo BYOK; 60-day free trial | Claude subscription or API | Free, “no paid version” | Free; optional paid models (Go $10/mo, Zen) |
| Platforms | macOS, Windows, Linux | macOS, Windows, Linux | Mac, Windows, Linux | macOS, Windows, Linux |
- Claude Code — code.claude.com/docs/en/overview · code.claude.com/docs/en/mcp · code.claude.com/docs/en/channels · github.com/anthropics/claude-code/blob/main/LICENSE.md
- OpenClaw — openclaw.ai · github.com/openclaw/openclaw · docs.openclaw.ai/cli/mcp · docs.openclaw.ai/tools/exec-approvals · docs.openclaw.ai/channels
- OpenCode — opencode.ai · github.com/anomalyco/opencode · opencode.ai/docs/skills/ · opencode.ai/docs/mcp-servers/ · opencode.ai/docs/permissions/ · opencode.ai/docs/rules/ · opencode.ai/docs/agents/ · opencode.ai/go
Frequently asked questions about AI harnesses
What exactly is an AI harness?
An AI harness is the software that wraps a language model so it can do work instead of only answer: a loop that keeps it going, tools it can call, memory that persists, and rules about what it may touch. The model reasons; the harness acts. Vodou is a harness that runs on your machine across any model provider, unlimited skills and unlimited MCP servers, with one memory underneath all of them.
Which is the best AI harness?
It depends on the job. For a codebase, Claude Code and OpenCode are the most mature harnesses today; for a chat assistant, OpenClaw has the most built-in channels. For one person using many models and tools, Vodou is built to be the harness above them: local memory that follows you into ChatGPT, Claude and Cursor, routing before inference, and tools that run together. Vodou is a public alpha — try it free for 60 days.
Is Claude Code a harness?
Yes. Claude Code is Anthropic's harness around Claude: it runs the model in a loop, reads and edits files, calls MCP tools, keeps CLAUDE.md memory and enforces permission rules — Wikipedia lists it as a canonical example. Vodou works alongside it: four hooks give Claude Code Vodou's memory on session start and every prompt, and Vodou's harness routes the same memory to every other tool you use.
What is the difference between an AI harness and an AI agent?
An agent is a model plus a harness. The model supplies reasoning; the harness supplies the loop, tools, memory and rules that turn reasoning into action. When people say "an AI agent booked my flight", the harness did the booking. Vodou is the harness: it hosts the memory, every MCP server you connect and every skill you install, and any model can be the reasoning inside it.
What is an AI harness vs a wrapper?
A wrapper is a thin interface around one model — the prompt is the product. A harness is a router with memory: it chooses among models and tools, keeps what it learns, and gets better with use, as Greg Isenberg put it, "a wrapper was one model doing everything and a harness is a router." Vodou is a harness in exactly that sense: swap GPT for Claude for a local Qwen model and the memory, skills and tools stay.
Is there an open source AI harness?
Yes — OpenClaw and OpenCode are MIT-licensed, and LangChain's Deep Agents is an open library. Vodou is open-core: the gateway, every bundled skill, the Vodou Bridge extension and the MCP servers are Apache-2.0 on GitHub (VodouAI/OS); the Rust engine that does routing and memory is proprietary and free to use with an account. You can read everything that touches your data.
What is the best AI harness for local LLMs?
A harness for local models needs to run locally itself. Vodou runs on your machine, binds only to localhost, and treats Ollama, LM Studio and a bundled llama.cpp as first-class providers next to every major hosted one, so a local model gets the same memory, skills and MCP servers as GPT or Claude. OpenClaw and OpenCode also accept local models; neither carries memory into the browser AI sites you use.
How can I build my own AI harness?
You need four parts: a loop, tool access (usually MCP), a memory store, and rules for approvals. LangChain Deep Agents and the Claude Agent SDK give developers those parts as libraries. If you want the result rather than the project, Vodou is a finished harness: install it on macOS, Windows or Linux, connect your models and tools, and extend it with SKILL.md skills instead of code.
Building on top of a harness? See the developer view: Vodou as an agent harness.
More from Vodou
- AI memoryMemory you own that compounds with every session.
- 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.
- Local AI agentRuntime, memory and tools on your machine.
- Claude Code memoryMemory in Claude Code, Cursor and Codex on every prompt.
One harness. Every model. Every tool.
Install Vodou, connect the AI tools you already use, and stop being the harness yourself. Free for 60 days, no credit card.
Updated