Connectors

A connector is access to an external system, granted to an agent. You add it once in Agents → Connectors, attach it to the agents that need it, and during their tasks those agents get that system's tools.

Without a connector an agent works inside Bounc only. With one it can post to Slack, read an issue in Linear, or call your own API — exactly the systems you allowed, for exactly the agents you gave them to.

What you can connect

Slack and Linear connect through the provider: you approve the access in its own window. Slack takes a user token only (it starts with xoxp-); its MCP server does not accept bot tokens. GitHub takes a personal access token. A Telegram bot takes the token from @BotFather: Telegram runs no MCP server of its own, so Bounc runs one in a container of its own that holds the token and nothing else.

Web search

Agents have no built-in search: an agent searches the web only when it is given a search connector.

• Brave Search takes a Brave API key. Bounc runs Brave's official server in a container of its own, which reaches Brave's API and nothing else.
• Tavily takes a Tavily API key, which goes to Tavily's MCP server only.
• SearXNG takes the address of your SearXNG server, which must have the json format on. The connector's container reaches that address alone. When the server is on an internal network, the engineer adds its address to the console's «Internal addresses».

Custom env secrets connectors from earlier versions are removed on upgrade: agents reach external systems only through MCP servers, and the agent's sandbox holds no secrets.

Custom MCP (JSON)

Paste one server's configuration in the form other MCP clients take:

{
  "mcpServers": {
    "everything": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-everything@2026.8.31"],
      "env": { "API_KEY": "…" }
    }
  }
}

• One server per connector. The command is npx, uvx, node or python. Pin the package's version (package@1.2.3 for npx, package==1.2.3 for uvx): without it every start installs the latest one, and the form says so.
• env values are secrets. They are stored encrypted, never shown after saving, and reach only the server's container, never the agent.
• Where the server goes. List the hosts the server needs over HTTPS: api.example.com or *.example.com. Besides them the container reaches only the npm and PyPI registries, to install its package.
• A remote server ("url" instead of "command") is saved as a Custom remote MCP connector.

The server starts once the connector is attached to at least one agent. The list and the connector's settings show its state: Starting, Running, or Error with the last lines of its log, secrets removed. Restart starts its container afresh.

Who can see it

Visibility is chosen at creation and cannot be changed afterwards:

• Workspace — available to everyone, attachable to any agent.
• Private — visible only to you, and attachable only to your own private agents: a workspace-level agent will not take a personal connector. It cannot be given to somebody else's agent either, not even by an admin.

Secrets

• Values are never read back. A stored token shows as stored, not as text. Leave the field empty and the stored value is kept; type a new one and it replaces the old one entirely.
• The agent sees tools, not a secret. Bounc's gateway holds the token and adds it to the requests to the MCP server; it never enters the agent's environment. A Telegram bot's token lives only in its MCP server's container, and that container reaches api.telegram.org and nothing else. A Custom MCP (JSON) server works the same way: its env lives only in its container.

Attaching to an agent

Open the agent, go to the Settings tab, Connectors section, and press Attach connector (the Connectors line in the agent's properties leads there too). One connector can go to several agents; its card shows how many hold it.

Detaching takes the access from that one agent. Deleting a connector revokes it from all of them at once — their running tasks lose the access immediately, which is why the delete dialog names how many agents are affected.

Next

• Agents — who connectors are given to.
• Credentials — secrets for workflow nodes, not for agents.