Tools & integrations

Every agent runs with a toolset that the daemon assembles for it. Some tools are built in; others — a real browser, and any MCP server you connect — are opt-in capabilities you enable once and every agent can then use.

The built-in tools

By default an agent gets:

• coding tools — read, write, and edit files in its group’s filesystem root,
• web search and a hardened web fetch (SSRF-guarded),
• mailbox tools — send_message, read_inbox, wait_for_reply,
• memory tools — search and write the group’s shared memory,
• home-file tools — read and update its own IDENTITY.md and notes.

Because a group’s root can be a symlink to a real project (see Core concepts), agents work on actual code rather than a throwaway sandbox.

Browser automation

Agents can drive a real Playwright browser. Perception is accessibility-tree-first: the agent reads an aria snapshot with stable element references and acts on those references, so it doesn’t depend on a vision model to see the page.

The suite covers navigation and interaction — browser_navigate, browser_snapshot, browser_click, browser_type, browser_fill_form, browser_select, browser_hover, browser_press_key, browser_go_back, browser_tabs, browser_take_screenshot, browser_console, and browser_network.

Install the browser binary once after upgrading:

pnpm exec playwright install chromium

Then enable browser automation on config → Browser Automation (or via environment variables).

Caution

A network-layer SSRF guard blocks loopback and private-network targets by default, so an agent can’t browse localhost or your LAN. For local development you can override it with the BROWSER_ALLOW_PRIVATE_NETWORK environment variable.

MCP servers

The daemon is a Model Context Protocol client. Connect an MCP server and its tools are discovered automatically and injected into every agent’s turn, namespaced as mcp__<server>__<tool> so they never collide with built-ins.

Three transports are supported:

• stdio — the daemon launches the server as a local subprocess,
• Streamable-HTTP — a remote server over HTTP, with optional bearer auth,
• SSE — a server-sent-events endpoint, with optional bearer auth.

Manage servers from the CLI:

bazilion mcp add <name> …     # register a server
bazilion mcp list             # list configured servers
bazilion mcp show <name>      # inspect one server and its tools
bazilion mcp test <name>      # connect and verify it responds
bazilion mcp enable <name>
bazilion mcp disable <name>
bazilion mcp rm <name>

The same servers are managed visually on the /config/mcp page.

File handling

Files move both ways between you and your agents, across the web UI, Telegram, and the CLI.

Inbound. Attach a file and the daemon classifies it: images are passed to the model as vision input, while other files are stored where the agent can open them with its coding tools. From the CLI:

bazilion agent chat <agent-id> --image ./screenshot.png
bazilion agent chat <agent-id> --file ./report.pdf

Outbound. Agents send files back with the deliver_file tool (up to 25 MB per file). A delivered image appears as a standalone image block in the web chat and as a photo on Telegram; other files arrive as downloads.

Note

File handling, the browser, and MCP tools are all available the same way from the CLI, the HTTP API, and the web UI — they’re capabilities of the daemon, not of any one surface.