Buddvisor

Self-hosted multi-agent workspace

A project room for a team of specialists.

Buddvisor keeps a work session, its documents, its agents, and the conversation in one place. You decide who is in the room, which model each one uses, and which tools they may touch.

View the app on GitHub See how a session works

01

A session is the project.

Name it, come back to it, and the history is still there. Several people can log in; each person sees their own sessions.

Create the room

A work session holds the orchestrator, the agents you assign, and the documents those agents are allowed to use.

Choose who speaks

Mention an agent directly, let the orchestrator route the message, or turn on brainstorming so the team takes turns.

Keep the record

Chat, formal tasks, and scheduled jobs all land in the same session. You can import, export, and share a session.

02

The orchestrator decides how the team works.

Each session picks one mode. Both use the agents, tools, and documents you already assigned.

Router

The orchestrator chooses one agent, several agents, or answers itself. Each specialist keeps its own slice of the conversation: your messages, and that agent’s replies.

Orchestrator-led

The orchestrator stays in the lead. It writes a brief and delegates. The specialist sees the brief, its tools, and its documents — not the whole chat — and the orchestrator writes the final answer.

Brainstorming is a separate mode. Agents take turns on a prompt you start, and you can interject. Routing does not run on those turns.

03

You configure the specialists. They do not arrive pre-built.

An agent is a name, a role, a system prompt, and a provider. One session can mix Claude, OpenAI, Gemini, DeepSeek, Qwen, Granite, and a local Ollama model.

Tools are checkboxes and settings on that agent, in that session. A folder, a database, a terminal, or web search exists only where you turn it on.

  • Orchestrator provider, model, and initial context
  • Agents assigned to the session, each with its own prompt
  • Documents assigned per agent, searched with local embeddings
  • Tools enabled per agent, including the orchestrator
  • Scheduled jobs and an inbound webhook for that session

04

Tools stay inside the session.

Agents work with the files, data, and media you gave them. The terminal is a persistent shell in that agent’s workspace, on Linux and macOS.

Workspace
A private folder per agent. Create, edit, and read files there. Assigned documents appear as links inside it.
Terminal
Named bash sessions that stay open across turns: working directory, environment, logins, and background jobs. A filesystem jail hides the app tree and home directory when the OS allows it. Its log is separate from one-shot command history.
One-shot shell
Run a single command in the workspace and exit. Use the terminal when the process needs to stay alive.
SQLite
An isolated database per agent: tables, queries, and updates, stored apart from the application database.
Documents and media
Upload an archive and ground answers with local embeddings. Assigned images, audio, video, and PDFs can be read by the agent that owns them.
Memory
In-memory keys for one agent, a shared pool for the session, optional long-term memory, and a reader for archived chat.
Schedule and webhooks
Run a prompt or a script on a cron or interval. An external system can post into the session with a per-session secret, and that post follows the normal chat pipeline.
Artifacts
An assistant can return an HTML block. The chat renders it in a sandboxed frame, so a chart or a small tool can sit in the thread.

05

It runs on your machine.

Buddvisor is a Node.js application with SQLite, a browser interface, and local embeddings. Provider keys stay in your environment. There is no hosted Buddvisor account to sign up for.

This page is only the introduction. The application is the repository you clone, configure, and run yourself.

github.com/mekanixms/Buddvisor
  1. Clone the app and install dependencies. Node.js 18 or newer.
  2. Set the JWT secret, the encryption key, and the provider keys you actually use.
  3. Start the server and open it in a browser.
  4. Create agents, then a session, then assign tools and documents.
  5. Chat. Turn on brainstorming only when you want the team to take turns.