The platform

Capability · Tools

Any MCP tool. Default-deny from day one.

Connect any MCP-compatible tool — and every one of them onboards default-deny, with credentials from the write-only vault.

monopea operates your actual stack through the Model Context Protocol: CRM, inbox, calendar, code, data — anything that speaks MCP plugs in and becomes an ability the agent can compose into its plans.

Openness is exactly why the gate exists. A new tool arrives with no standing trust: its mutation-capable calls route to review from the first invocation, and calls the platform cannot classify fail closed. You extend the agent’s reach first and grant its autonomy second — deliberately, per tool.

What connecting a tool actually grants

Connecting an MCP tool registers its capabilities with your agent — it does not grant permission to use them freely. Reads can flow, but mutation-capable calls from a newly connected tool are held at the approval gate by default, so a fresh integration cannot take an outward action you have not seen.

From there, trust is granted the same way as everywhere else on the platform: explicit per-tool policies, auditable and revocable. The catalog can keep growing without your risk surface growing with it.

  • New tools start at default-deny — connecting is not authorising
  • Autonomy is granted explicitly, per tool and per operation
  • Grants are auditable and revocable at any time

One agent composing every tool

Because every integration feeds a single agent rather than a pile of separate automations, abilities combine: pull context from your documents, update the CRM, draft the email, schedule the follow-up — one plan across four tools, with each outward step passing through the same gate.

That is the practical difference between a tool catalog and a stack of integrations: composition. The agent treats tools as verbs it can sequence, and the governance layer treats them all identically.

Credentials come from the vault, not the prompt

Tools authenticate with secrets the model cannot read. A connection’s API key lives in the write-only vault and appears in tool calls as {{secret:NAME}} — substituted only at dispatch, scrubbed from results. Connecting a powerful tool never means pasting a powerful credential into a prompt.

Scoping completes the picture: each connection carries exactly the access you granted, and revoking a connection or deleting its secret shuts it off cleanly.

Bring your own tools

MCP is an open standard, so the catalog is not a fixed list. Internal tools you build, niche services, self-hosted systems — if it speaks MCP, it plugs into the same plans, memory, and gate as everything else, with no special-case integration work.

Homegrown tools get no homegrown trust, either: your own MCP server onboards default-deny exactly like a third-party one — which is precisely what you want when an agent is calling code you wrote on a Friday afternoon.

Why it matters

Reach without exposure

Every new tool starts gated. Connecting is safe; autonomy is a separate, explicit grant.

Your real stack

The agent works in the tools you already use — anything MCP-compatible plugs in.

Secrets stay in the vault

Tool credentials are referenced as {{secret:NAME}} and never shown to the model.

FAQ

MCP tool catalog, in short

Which tools can Monopea connect to?
Anything that speaks the Model Context Protocol — CRMs, inboxes, calendars, data stores, developer tools, and MCP servers you build yourself. The catalog is open by design.
Does connecting a tool let the agent use it immediately?
Reads can flow, but mutation-capable calls from a new tool are held at the approval gate by default. You grant autonomy per tool, explicitly, once you have seen it work.
What about calls the platform cannot classify?
They fail closed. If a call cannot be matched to a known, granted policy, it routes to review rather than dispatching.
How are tool credentials handled?
Through the write-only vault: stored envelope-encrypted, referenced as {{secret:NAME}} in calls, substituted only at dispatch, and scrubbed from tool results. The model never sees the value.