What is Vibemind?
Vibemind is an AI-native operating system for a fleet of agents. You type or message an intent — a deterministic router decides who does what, agents in a Rust runtime do the work, and two memories keep what was decided and what your material is. Your ideas appear as 3D bubbles you can navigate, connect, and transform.
It's open by design. The public repository, vibemind-space/vibemind-os, is on GitHub under MIT — yours to fork, audit, and self-host. Which components are open under which licence, and where the gaps still are, is spelled out on the licensing page.
Watch the vision — Vibemind in motion.
How it works
Five stages, each one observable. Deterministic in the hot path — no brittle LLM sitting between you and your work.
You send an intent
Type it, or message it from Telegram or WhatsApp. Voice is in progress.
The Brain classifies
Regex-matched against a 119-capability registry first — deterministic and auditable. The semantic fallback is not working yet.
The plan dispatches
The routing decision becomes concrete steps: a database operation, a service call, or a task handed to an agent.
An agent does the work
Agent tasks go through an order table to OpenFang, the Rust agent runtime, and come back with a result.
The result is checked
A validator checks the outcome against ground truth before it counts, and the decision is recorded for later.
See it live.
Three captures from our workbench. Click to play — no edits, no scripts, real agents on real workflows.
Idea Capture & Organization
Ideas captured, classified and linked. The formatting into kanban, mindmaps and SWOT seen in this capture was removed in June 2026.
Knowledge Graph
Every artifact auto-indexed with semantic search and cross-referencing.
Deep Research
🇩🇪 DEStructured web research saved directly to your knowledge graph.
Two graphs, two jobs
Most agent tools forget everything between tasks. Vibemind keeps two separate memories — one for how it decides, one for what your code is.
The routing graph
Learns how the system decides
Collections in Qdrant — episodic, semantic, decisions, a self-model — holding how intents got routed. Successful routes feed an offline learner that re-weights the matrix.
13
memory collections
The code graph
Knows what your code is
fungus-search indexes your codebase into semantic chunks. A full re-index took about 25 minutes on GPU, and the automatic updater is currently not running. Separate system, separate job.
67k
code chunks, last full index
The longer it runs, the better it routes. That compounding is the part competitors can't bolt on afterwards.
Your whole system, one message away
You don't have to be at the machine. Talk to the system from your phone.
- 01
You message
Telegram today, through OpenFang. WhatsApp runs through OpenClaw.
- 02
An agent picks it up
Today a single operations agent answers; routing through the Brain comes back with the Brain service.
- 03
Work happens
The agent runs in OpenFang with the tools it is allowed to use.
- 04
Answer comes back
On the same channel you sent it from.
The runtime has adapters for 42 channels; Telegram and WhatsApp are the two configured today. Every step leaves a trace you can read.
Who it's for
Founders & Product Leaders
- •Dictate product ideas on the go — auto-organized into a visual brainstorming canvas
- •Speak requirements that auto-flow into user stories and architecture diagrams
- •Generate meeting briefs from your knowledge graph with one sentence
- •Run competitive research and get structured comparison tables by voice
- •Draft stakeholder emails using context from your project data
Built with
Agent Runtime
Cognition
Memory
Interface
Infrastructure
Own your stack
Independence isn't a slogan here, it's literal. The same principle that runs the agents runs the infrastructure: nothing you couldn't rebuild yourself.
Your own data
Supabase, Postgres and the vector stores run on our own server, not in someone else’s cloud. Mail and model calls still use outside providers today.
Your own git
A self-hosted Gitea server ships in the stack, ready to host code and issues next to the open core.
Your own models
Adapters for 42 model providers, including local models through Ollama. Today most agents run on Claude; switching is configuration, not a rewrite.
Your own memory
The routing graph and the knowledge graph live with you, not with a third party you have to trust.
Built in the open
Open by design
There's already an option for closed, walled-garden AI agents: OpenAI, Anthropic, Microsoft Copilot. There hasn't been an open one. Vibemind wants to be the Linux of agentic AI: the umbrella repository is on GitHub under MIT, the runtime under Apache-2.0 or MIT, and a few components carry other licences, including non-commercial ones. The licensing page lists every component and its licence.
Read it, fork it, audit it. Self-hosting works today if you are ready to read configuration; it is not a one-command install yet. The repo is early and moving fast — read the code, open an issue, and the roadmap bends toward the people who show up.
Open source · Built in the open · No telemetry in the runtime
Join the beta.
The beta is for Vibemind as a whole — every space brought together in one system. We roll it out to this list first, once Vibemind is stable. Single spaces you can request today.
The umbrella repo is public under MIT. Beta access is first-come.
— Felix & Stéphane
Co-founders · Vibemind