Tools & MCP Access¶
Everything an agent does, it does through a tool — a single named action like reading a file, querying a data table, drafting an email, or looking up an invoice. Agents have no hidden powers: if there's no tool for it, an agent can't do it. This page explains the two kinds of tools BotBoss supports, how they're granted to an agent, how the same tools are reached from an external MCP client, and exactly what is and isn't possible.
The two kinds of tools¶
Built-in platform tools. These are the capabilities BotBoss provides directly —
managing tasks and workflows, querying and writing your data tables, building dashboards
and artifacts, working with files, using skills, reading and writing an agent's memory,
sending messages, and raising approvals. They're always part of the platform; an agent
just needs the matching permission to use each one. The full, always-current catalog —
every tool, its access level (read / write), and the exact permission it requires —
is the Agent tools reference.
Connector tools. These come from an integration you've connected — QuickBooks, Ramp, Plaid, Gmail, Google Drive, and the rest of the connector fleet. A connector's tools only exist for an agent once two things are true: the connector is connected at the workspace level, and it has been granted to that specific agent. Connect without grant means the agent still can't see it — access is opt-in per agent.
How tools are granted to an agent¶
An agent's toolset is the result of a few layers, each of which can only narrow what the agent can do — never widen it:
- The tools you pick at hire time. When you hire or edit an agent you choose its tools and skills. That's the starting set. See Agents.
- Permissions (scopes). Each tool requires a matching permission. An agent only holds the permissions you've given it, so a tool it lacks the permission for simply isn't available to it — it never even appears.
- Connector grants. A connector tool is available only after the connector is connected and granted to that agent. You can grant, revoke, disable, or reconnect a connector at any time from the Connectors page.
- It can never exceed the person it's acting for. When an agent acts on your behalf — for example when you ask it something in chat or over Slack — its reach is the overlap of what the agent is allowed to do and what you are allowed to do. An agent can never reach data or actions you couldn't reach yourself.
- Approvals for sensitive actions. Any given tool can be set to run automatically, to ask you once per task, or to ask you every single time. Actions that move money or send things on your behalf are held to a confirmation step by default — the agent proposes, and a person signs off before it happens.
Because of this layering, what an agent can actually do at any moment is the intersection of all of them. Tighten any layer — remove a tool, narrow a permission, revoke a connector — and the agent's reach shrinks immediately.
Reaching the same tools over MCP¶
BotBoss also exposes these tools over the Model Context Protocol (MCP), so an external client — for example an MCP-capable desktop assistant — can connect to your workspace and use the same catalog. The rule that keeps this safe is simple: an MCP session acts with the permissions of the person who authorized it. A teammate connecting over MCP sees and can do exactly what they could do in the app — no more. The same permission checks, connector grants, and approval steps apply identically whether a request comes from the app, from Slack, or from an MCP client.
For safety, BotBoss connects only to a curated set of vendor integrations. You can't point an agent at an arbitrary, self-supplied MCP endpoint — connectors are ones BotBoss has vetted and governs.
What's possible — and what isn't¶
Possible, once connected and granted, and within the acting person's own access:
- Query and write your data tables; build dashboards and artifacts.
- Create, assign, run, and track tasks and workflows.
- Read and organize files; discover, follow, and share skills.
- Read and search email and draft replies; browse and read files in connected drives.
- Read your accounting, card-spend, and banking data; prepare draft financial entries.
Not possible, by design:
- No hidden actions. If there's no named tool, the agent can't do it.
- No exceeding your access. An agent acting for you can never see or change something you couldn't yourself.
- Connected never means granted. A connected integration stays invisible to an agent until you grant it to that agent.
- Email is read-and-draft by default. Agents don't autonomously send email unless sending is explicitly enabled for that connector.
- Read-only financial connectors stay read-only. For example, BotBoss reads your QuickBooks and Stripe data but never creates charges, refunds, or postings back in those systems on its own — the limit is enforced in code, not just by convention.
- Money moves and outbound sends need a human. Actions that spend, transfer, or send on your behalf are gated behind an explicit confirmation.
- No arbitrary MCP endpoints. Only curated, vetted connectors — never a self-supplied server URL.
Staying current¶
The Agent tools reference is generated directly from the platform's code and kept true by an automated check that fails the build if a tool changes but the reference wasn't regenerated. So the list of tools and the permission each one requires is always in lockstep with what the platform actually does — it can't quietly drift out of date.
See also¶
- Agents — hiring an agent and choosing its tools and skills.
- Connectors & Integrations — connecting the systems your agents use.
- Gmail, Drive & QuickBooks — connecting and granting these three, and their exact limits.
- Sharing & Permissions — how sharing and access roles work across every resource.
- Agent tools reference — the always-current catalog of every built-in tool and its required permission.