Profiles become Agents
A Profile is an Agent template: SOUL.md, IDENTITY.md, AGENTS.md, TOOLS.md, a default model and skills. Spawning one creates an Agent with its own documents and conversations.
Bazilion documentation
Bazilion is a local multi-agent runtime for one person. Agents run on
Pi’s coding agent with the LLM
provider of your choice: a ChatGPT login, an API key for OpenAI, Anthropic, Google and many
others, or a local LM Studio, Ollama or llama.cpp. They work in Teams that share a workspace,
memory and a mailbox, run on schedules, and are reachable from a web UI, a CLI and Telegram. One
daemon on your machine owns all state in ~/.bazilion.
It is not a hosted platform: it is a daemon, a CLI and a web UI you run yourself, under the MIT license.
Bazilion 0.26.0 simplifies the product for a single local operator. It needs Node.js
^26.12.1, and a home from an earlier release is refused: back it up and start fresh, as
described in Upgrade to 0.26.0. Read
what changed. Linux is the tested platform; macOS runs commands
on the host and is untested.
Getting started Install, open the dashboard and spawn your first Agent.
Core concepts Profiles, Agents, Teams, policy, memory, the mailbox and triggers.
Tools & integrations Web research, the managed browser, MCP, image generation and files.
Configuration Providers, settings, secrets, tokens and the on-disk layout.
Working with Agents
Section titled “Working with Agents”- Coding with Agents: the shell, the Docker sandbox and repository instructions.
- Reviewing what changed: live tool output and Git change review.
- Saved results and conversations.
- Follow-up queue and Agent questions.
- Telegram and Attention notifications.
Running your installation
Section titled “Running your installation”- Daily operations: diagnose a silent Agent, schedules, Attention and reviewed learning.
- Growth and retention: what grows in a home and what to delete.
- Backup and recovery: back up, restore and reset.
- Private access: reach the web UI from other devices over Tailscale.
How the pieces fit
Section titled “How the pieces fit”Pi runs the agent loop
Pi owns the turn loop, provider calls, compaction and the JSONL transcript. Bazilion adds Agents, Teams, credentials, tools and the daemon around it.
Teams share a workspace
A Team is a directory (or a link to your project), a USER.md about you, shared memory and its members. Team templates spawn a whole Team at once.
Policy says who may message whom
An open Team lets members message each other freely; a restricted one needs an allow edge. You can always chat with every Agent.
A mailbox between them
Agents message each other with send_message, read_inbox and wait_for_reply; an idle
Agent with unread mail wakes up.
Shared memory, indexed
Each Team keeps Markdown notes in <team>/memory/, searched with qmd (BM25), so decisions
outlive the chat that produced them.