Give AI a place.
Vancetope gives AI a place — and with it, an application: a workspace of documents, sheets, canvas and boards where agents sit with your data and don’t chat on the side, they do the work. For an answer in five minutes or an assignment that runs for days. For work and for life, kept apart. On your own server.
Self-hosted with Docker in a few minutes — bring your own model. Install →
- Self-hosted
- Bring your own model
- MCP both ways
- Web · Desktop · CLI · Mobile
- No telemetry
What Vancetope is
Where AI lives — and new AI grows.
A place for AI
Not a text box, but somewhere to live: right next to your data, with documents, apps, memory and tools. With a place, AI finally has something real to do.
Where new AI is made
A workshop for testing new kinds of agent loops: 18 engines, each a different way to think — deep task trees, bulk workers, fast decision models — and Nutrimat, a lab for experimental loops. Because the loop lives in code, not in a prompt, a new idea is a new engine — and it runs here first, beside real work.
One whole system
Not a chat with plugins bolted on. Apps, agents, memory, triggers, clients and security are one system built on one idea — everything is a document — so every part knows every other.
Real surfaces
Real apps — AI-made and connected.
Not a chat window, but the tools you actually work in — all AI-native: have them created, filled, reshaped and linked. Because everything is a document, it all hangs together, and it’s shared live — people and agents at once.
Cortex
Chat, document and execute on one surface — every kind of document, with the agent right beside it.
Workbook
Pages in a block editor — with forms, quizzes and runnable script blocks.
Sheet
Spreadsheets with functions — pipe the results onward.
Canvas
Think spatially: nodes, links and groups on an open plane.
Kanban, Issues & GTD
Plan work, track it through, keep your own list clean.
Scribble & Designer
Handwriting the agent can read and draw on; design collections with a live preview.
See it in action
What people get done with it.
Real walk-throughs — one request grows into research, charts, documents and decks. Each is a slideshow you can follow step by step.
Research a topic, end to end
Deep-research with Marvin, reshape it into a mindmap, illustrate it, compile a document — and export a PDF. All in one conversation.
Size up a stock with live data
Pull a stock’s prices off the web, chart them as candlesticks, and get an assessment that cites the numbers it stands on.
Stress-test an idea with a panel
Stand up a reusable council of AI personas, throw a pitch at it, and get one synthesised verdict — the disagreement kept intact.
…and more — data analysis, research to slides. See all how-tos and stories →
What makes it tick
A few ideas do most of the work.
Everything else is a consequence of these.
Everything is a document
Recipes, prompts, skills, schedulers, hooks, guards, settings, even apps — all documents. A new automation is a new document. The system reshapes itself from its own data, and so can agents.
Agents drive (almost) everything
Nearly every capability is a tool: write documents, spawn workers, set triggers, research, delegate. You give the direction and step in whenever you want — the flow stays visible.
A fat harness, not a thin one
Most agent tools are thin harnesses: a slim loop around the model, and the model decides everything. Vancetope is a fat harness — engines are Java algorithms with a lifecycle, and the code decides when the model is called, which tools it gets and when work is done. Few engines, many recipes. That’s what lets an assignment run for days without drifting.
Built to work in, together
Cortex unites chat, documents and execute on one surface — live-edited with other people and with agents in the same session. Presence, merge, versioning and discussions included.
Projects draw the boundaries
A project is a bounded area with its own documents, config and agents. Settings, memory and permissions inherit down a scope cascade — tenant, project, user, session.
It runs itself
Schedulers, lifecycle hooks and webhooks fire recipes, scripts or workflows. Persistent agents keep a journal, wake on their own schedule — and pause to ask instead of guessing.
Go deeper
There’s a lot under the surface.
Vancetope has grown well past a chat with documents. Four tours, each one page:
Everything you work in
Fifteen apps and thirty-odd document kinds — from workbook and kanban to handwriting, audio and apps the agent builds itself, sandboxed.
agents & enginesWork that doesn’t fit into a turn
Deep research trees, bulk runs over thousands of records, fast decision models, councils, and persistent agents with a home, a journal and a schedule.
clients & integrationsOne brain, many doors
Web, desktop, terminal, mobile, a browser extension — and MCP in both directions, guarded by a real OAuth server, so other agents can work with yours.
security & self-hostingYour server, your keys
End-to-end encrypted documents, sandboxes on every machine an agent touches, guards before dangerous commands, quotas — and nothing phoning home.
Work and life
One place for whatever you’re working on.
Work or personal — Vancetope holds both, kept apart in separate tenants. A five-minute question or a project that runs for weeks; something off the web, a document to reshape, files right there on your machine. It’s open-ended on purpose: the stuff you’d otherwise scatter across a dozen apps tends to just fit here.
Project kits
Start a project ready to work.
A kit is a git bundle of skills, recipes, tools and settings. Install one and a project arrives preconfigured — thinking patterns, worker recipes, tool packs, defaults. And kits inherit, so they compose:
The interesting part isn’t the bundled ones — it’s that a kit is just a git repo. Point a project at yours and it’s productive from the first message.
The proof is the product
Written by agents. On purpose.
Every line of Vancetope is AI-written — directed, reviewed and shaped by one human. That’s not a confession, it’s the whole point: Vancetope is what working with agents over months actually builds — a large, coherent system you can run, read and change. The tool and the demo are the same thing.
Questions
Asked often enough to answer here.
Is Vancetope a product I can buy?
No. It’s a personal project with public source — no sales, no support contract, no hosted service. You run it yourself, on a laptop, a server or a cluster.
Which AI model does it use?
Yours. Vancetope ships no model: connect a hosted API or a local one — any OpenAI-compatible endpoint works. Which models hold up under agentic load is measured in the benchmarks.
What do I need to run it?
A machine with Docker. Two commands install the stack and set up your first tenant and user — a few minutes. Get started.
Does it send my data anywhere?
Only to the model endpoint you configured, and only what the agent sends it. No telemetry, no licence check, no account with us. Documents can be encrypted so that even your own server can’t read them. Security & self-hosting.
Is it open source?
Yes — GNU AGPLv3. Use it, run it, change it and share it; if you offer a modified version to others over a network, you share its source too. The licence text is on GitHub.
Can my own coding agent work with it?
Yes. Vancetope is an MCP server with three doors — tenant, project and agents — behind OAuth 2.1, so Claude Code, Cursor and others connect with a browser login. MCP doors.
Honest bit
It’s a personal project.
I built Vancetope to develop LLM agents and shape them until I could actually work productively with them. Over time everything I found interesting went in — plus a few things I think others might get something out of.
No support contract and no roadmap promises — it’s a project, not a product. The source is open under the GNU AGPLv3: read it, run it, change it, take it for a spin.
— Mike Hummel