VibeXP is one shared knowledge base for your prompts, rules, memory and past work. Claude Code, Cursor, ChatGPT and every other MCP client read from it before a task and write back to it after, so each session starts from what the last one learned, for you and for everyone on your team. It is free, open source, and it runs on your own server.
The instruction that finally worked is buried in a chat history, or in a teammate's notes you will never see. So everybody reinvents it, slightly worse, every time.
One prompt library, every toolEvery session starts cold. You describe the project, the stack and the standards again, to every tool, and nothing your AI worked out yesterday is there today.
Memory it reads back firstRules in CLAUDE.md here and .cursorrules there, outputs in scratch files, lessons on one laptop. None of it reaches the next session, let alone your team.
One shared, searchable homeYour AI searches the relevant prompts, rules, memory and past work first, so it starts with everything already learned instead of a blank slate.
It saves new lessons, updates memory and stores what it produced, through the MCP write tools. An explicit step the agent takes, not a guess.
That richer knowledge is waiting, in every tool, for you and for every teammate. Built-in AI memory is one tool and one person. This is all of them.
vibexp_io_render_prompt(slug: "code-review")Give every tool the same rulesvibexp_io_get_resource(resource_type: "blueprint", slug: "claude-md")Remember what it learnedvibexp_io_create_memory(content: "pool_max 50 fixed it")Find it again by meaningvibexp_io_search(query: "why do requests time out?")Keep the report it wrotevibexp_io_create_artifact(title: "Q3 incident review")Follow an agent while it worksvibexp_io_post_to_feed(content: "migration 3 of 5 done")Connect how things relatevibexp_io_link_resources(relation: "governed-by")Alsocomments·metadata filters·access analytics·GitHub import·every feature
Put your prompts, rules and memory in one place today. Your own AI gets more consistent from the first session, because it reads what you saved before it starts.
Invite people into a team workspace and it compounds. One person's prompt, rule or hard-won lesson makes everyone's AI better, instead of each person quietly reinventing it.
The bundled file runs the published image next to PostgreSQL with pgvector. The app and its API share one port, and local evaluation signs you in without an identity provider, so the first thing you see is the app itself.
git clone https://github.com/vibexp/vibexp.git && cd vibexp && docker compose up -d
Then open http://localhost:8080. Data lives in the pgdata volume, so it survives a restart.
Point any MCP client at your instance. The first call opens a consent screen in the browser; approve it once and the client keeps its token. Cursor, VS Code, Gemini CLI, Codex and ChatGPT take the same URL.
claude mcp add --transport http vibexp http://localhost:8080/mcp/v1/common
docker run -p 8080:8080 \ -e DB_HOST=your-db-host -e DB_PASSWORD=secret \ -e ENCRYPTION_KEY="$(openssl rand -base64 24 | cut -c1-32)" \ -e FRONTEND_BASE_URL=https://vibexp.example.com \ ghcr.io/vibexp/vibexp:0.18.0
Postgres must have the pgvector extension available. The image reads every setting from environment variables.
FRONTEND_BASE_URL=https://vibexp.example.com
ENCRYPTION_KEY=$(openssl rand -base64 24 | cut -c1-32) SESSION_ENCRYPTION_KEY=$(openssl rand -hex 32) DB_PASSWORD=...
INSTANCE_ADMIN_EMAILS=you@example.com docker compose logs app | grep "SETUP URL"
Leave FRONTEND_BASE_URL at localhost on a public host and anyone can sign in as any user.
Imageghcr.io/vibexp/vibexp:0.18.0·Release notes ↗·Upgrading·Kubernetes and bare binary
Built-in memory lives inside one tool and one account. VibeXP is shared across every tool you use and everyone on your team, so your context follows you from Claude Code to Cursor to ChatGPT, and your team's knowledge compounds instead of being locked inside one vendor's app.
Memory ↗Anything that speaks the Model Context Protocol and its OAuth 2.1 sign-in flow: Claude Code, Cursor, VS Code, Gemini CLI, Codex, ChatGPT and more. Every one of them connects to the same single endpoint on your instance, and one connection reaches your prompts, rules, memory, artifacts and feeds.
MCP server ↗No. You paste one URL into your MCP client, it opens a consent screen in the browser, and you approve it once. There is no key, client id or secret to copy. API keys do exist, separately, for the CLI, the REST API and your own scripts.
How OAuth connect works ↗You run it. VibeXP is free and open source under the MIT license, and ships as one container image that serves the app, the API and the MCP endpoint from a single port, next to PostgreSQL with pgvector. docker compose up gets you a local instance with nothing to configure; a public one needs a URL, a few secrets and an identity provider.
Self-hosting ↗Yes. Centralize your prompts, rules and memory today and your own AI gets more consistent from the first session, because it reads what you saved before it starts. Invite your team whenever you are ready and the same knowledge base becomes shared.
Quick start ↗Your tools write what they learn back into VibeXP over MCP, then read it before the next task. It is an explicit step the agent takes through the memory tools, not a guess, so the knowledge base grows with every session and travels with you across every tool.
Memory ↗On the server you run it on, in your own PostgreSQL database. There is no VibeXP cloud behind the open-source build for your data to pass through. Semantic search, email and file attachments are opt-in, and each one talks only to the provider you configure for it.
Self-hosting ↗Stars are how the next person tired of re-explaining their project to an AI finds this. It is the whole of the distribution: no ad budget, no sales team, no hosted plan to upsell.
One command to run it, one to connect your AI, and the next session starts from everything the last one learned, for you and your team.