Connectors & MCP reference

// plug the station into your tools — and into your other harnesses

Built-in tools

Agents ship with a broad real-tool belt. Everything an agent can reach outside its own reasoning belongs to one of eight capability families, and every family is gated by the station's capability model (object = capability) — you grant it by placing the gear, and you can revoke it the same way.

Family	What it grants
Web & browser	Search the web, fetch pages, and drive a real browser — navigate, read, click, type.
File cabinet	Read, search, and write files in the agent's workspace.
Workbench	Run shell commands and verify code — the code-execution bench.
Task delegation	Delegate subtasks to summoned crew, summon new agents, and schedule routines.
Comms relay	Send a message out to a connected chat — only chats someone already opened with this station.
Media studio	Generate images from text and analyse images, on the key the agent already uses.
Memory notebook	The agent's private memory, task plans, and saved skills.
Jukebox	Search and control your Spotify — play, pause, queue, "what's playing".

Browsing that actually browses

The web family is not just a search box. An agent can run a search, fetch a page, and then drive a real browser against it — navigating, reading the rendered page, clicking, and typing. When a page is unreachable or throws up a challenge, the agent reports that as the result rather than inventing a plausible summary of a page it never loaded; a 403 or a 404 is an answer, not a hidden failure.

When a site genuinely requires a login, the agent can't and won't ask you for the password. It asks the station to open a visible browser window and you sign in yourself — your password is typed into real Chrome and never passes through the agent or the model. The resulting cookies live in the station's browser profile, so later runs stay signed in without ever having handled your credentials.

MCP connectors

bdibranding speaks Model Context Protocol. A curated connector catalog covers popular services with one-click setup; connectors that need auth use OAuth 2.1 dynamic client registration or a pasted key. Any third-party remote http(s) MCP server can be added by URL — your agents' reach grows with the ecosystem, not with our release schedule.

Local stdio servers (npx …, uvx …) are supported only through an agent-owned SAFE CELL. bdibranding probes Docker before connecting, runs the server inside that isolated cell, keeps connector secrets out of process arguments, and never falls back to spawning the server directly on your host. If no real Safe Cell agent exists, the local-server route is unavailable rather than pretending the process is isolated. Safe Cells have no internet by default; turn on SAFE CELL internet access in Settings › PERMISSIONS for servers that need to download or reach an online API.

Connecting a specific service? CONNECT A PLATFORM walks the whole decision end to end — which of the two routes applies, how a key is stored and substituted without the model ever seeing it, unattended use, and what each failure message means.

Messaging channels

Connect Telegram, Discord, Slack, Matrix, or Signal and your agents can report to you — and take tasking from you — where you already are. Scheduled and Night Shift runs can deliver their results straight to a channel. Matrix talks to whatever homeserver you point it at, and Signal talks to a signal-cli bridge you run yourself; bdibranding never picks the server for you.

Outbound messages are a capability like any other: an agent can only write to a chat someone has already opened with this station, never to an arbitrary address it decides on.

Telegram as a remote control

Telegram goes further than notifications. From your phone you can hold a real threaded conversation with a specific agent, start and steer runs, answer an approval an agent is waiting on, and orchestrate the crew — the same command surface as COMMS, reachable when you're nowhere near the machine. The station is still doing the work locally; Telegram is just the window you're looking through.

As the paired owner you can also stop all station work from the phone: send /halt, then /halt confirm within 60 seconds (/stop ends only that chat's run). RESUME AUTOMATION is only in the station window.

Use bdibranding from your editor or another app

Already run a code editor with ACP support, Claude Code or another agent harness that speaks MCP, or a script that talks to an OpenAI-style API? They can drive your station, with a separate, limited key, once you allow it. bdibranding agents then sit alongside the tools you already use instead of replacing them.

Turn it on. Open SETTINGS › PERMISSIONS and find EDITORS, MCP CLIENTS AND /v1 TOOLS. It is off by default; turning it on asks twice, then makes a new key.
Copy the port and the key. The same section shows the station's port and the key, with COPY PORT and COPY KEY. The desktop app picks a new port every launch, so the editor and MCP bridges that come with the app find the running station and its key on their own; the section shows where they are installed. A /v1 tool uses http://127.0.0.1:PORT/v1 with the key as its bearer token.
Point your editor or app at it and start a run from there.

What the key can do. It starts runs on your own provider keys and reads their results. It can never approve or decline anything, turn on Full Access, change a setting or see your saved keys: those stay in the bdibranding window. A run it starts never gets Full Power or FULL BYPASS and uses only the gear its agent has. Each change such a run wants to make (a file, a command, a new website) asks you in the bdibranding window, on an approval card that names the app, and is declined if nobody answers in time. /v1 runs answer in text and are given no tools.

What it can read. By default only the runs and conversations it started. INTEGRATIONS CAN READ ALL MY CONVERSATIONS (off by default, asks twice) widens that to everything on the station.

Taking it back. MAKE A NEW KEY replaces the key, and TURN OFF retires it; either way the old key stops working for every new request. Turning it back on makes a new key, so give it to each editor again.

Local by default

Every connector and channel is opt-in. The default station touches the network only for the model provider you configured and, at startup, two public model and price lists that carry nothing of yours (see the privacy page); anything wider is a grant you made on purpose and can revoke in Settings.