Let an agent drive it —
One MCP URL gives Claude or Cursor the same 459 tools the dashboard uses — OAuth 2.1, database-enforced permissions, and a confirmation on every call that changes something.
Agent access usually means
Handing an LLM a full-access token is how a "summarize my traffic" request turns into an agent that can silently wipe a device's rules.
For teams wiring an AI coding agent, a support bot, or their own automation into Busymate — querying captures, controlling devices, or driving a browser — without handing it unscoped, unaudited power.
Busymate's MCP server exposes the same 459 tools the dashboard itself uses, authenticated with OAuth 2.1 (dynamic client registration plus PKCE — no pasted keys), scoped to the caller's role by Postgres row-level security, and gated so every destructive tool demands an explicit confirmation. Nothing an agent does is invisible: every call lands in the audit trail.
Real power,
The same identity authenticates MCP, REST, WebSockets and BusyBro — one token, enforced everywhere.
OAuth 2.1, done right
Dynamic client registration plus PKCE — no pasted API keys. The token your agent gets also authenticates REST, WebSockets and Edge Functions.
Confirm-gated, database-scoped
Every call is scoped to the caller's role by the database itself, and destructive tools demand an explicit confirm before they run.
Browser automation, not just data
The bmc CLI connector lets an agent drive named Chrome profiles over the DevTools Protocol — open a URL, run a check, mock a response — with capture watching the whole session.
- MCP
- BusyBro
- CLI
- Dashboard
From one URL to
Standard MCP the whole way — nothing proprietary to learn.
- 01
Add the server
Point Claude Code, Cursor, or any MCP client at the one connect URL — one click for supported clients, or paste the config.
- 02
Authorize once
Approve the OAuth 2.1 flow in your browser. The agent gets a role-scoped token — no key to copy or rotate by hand.
- 03
Ask, then confirm
Ask it to query captures, control a device, or run a browser check. A read happens immediately; a write pauses for your explicit confirm.
How it works
Standard MCP — the client does the talking, the server enforces the limits.
- 01
Connect the server
Add the one MCP URL to Claude Code, Cursor, VS Code or any MCP client — a one-click link for the clients that support it, or a pasted config from the install page. The client registers itself with the server on first contact; there is no API key to create, copy into a settings file, or rotate later.
- 02
Authorize in the browser
The first call opens a browser sign-in. You approve the client once and it receives a token scoped to your role — the same identity that authenticates the REST API, the WebSocket stream and BusyBro, so a permission you grant or remove in the dashboard applies to the agent on the next call, with nothing to redeploy.
- 03
Ask, and confirm what changes
Ask the agent to summarize a device's traffic, replay a failed request or run a browser check. Reads happen immediately; any tool that changes something pauses and shows you exactly what it is about to do until you confirm. Every call, read or write, lands in the audit trail.
Before you connect an agent,
Which clients work?
Any MCP client: Claude Code, Claude Desktop, Cursor, VS Code, ChatGPT, and your own agents. The install page has copy-paste configs and one-click links for the clients that support them.
Can the agent do more than my account can?
No. The token is scoped to your role and the database enforces it, so the agent sees and changes exactly what you could — no more, on any surface.
What stops an agent from wiping something?
Destructive tools are confirm-gated: the call pauses and shows the full payload before anything runs. Nothing an agent does is invisible — every call is written to the append-only audit trail.
Is there a hosted assistant, or only bring-your-own?
Both. BusyBro is the built-in teammate that already knows your traffic and devices — in the dashboard, on Telegram and over MCP; your own agent gets the same tools through the same server.
Connect your agent to
One click for Cursor or VS Code, or paste the config — authorize once, then call any tool.