Skip to main content
Tools are how an agent reaches the rest of your company, and it uses them two ways. To find out. It searches Slack, Linear, Drive, or any of 760+ apps at the moment it needs the answer, rather than working from whatever you pasted into the instructions. To act. It raises the invoice, drafts the document, updates the record. Attach none and it can only talk.

Piece tool

A single action from one of your connected apps.

Flow tool

One of your own flows, handed over as a tool.

MCP server

An external server, for systems that aren’t pieces.

​
Piece tools

You attach individual actions, not whole apps. An agent given Send message can post to Slack and nothing else. Each input on that action is one of three modes, so you pin what has to be exact and leave the rest to the agent.

Let agent decide

It writes the value.

Set value myself

You pin it.

Leave empty

Nothing is sent.
Pin the Slack channel, let the agent write the message.

​
Flow tools

The most underused one. When part of a job has to happen the same way every time, build that part as a flow and hand it over. The agent chooses the moment. The flow guarantees the steps.

​
MCP servers

Paste the URL and pick how it authenticates: none, an access token, an API key, or custom headers. Streamable HTTP, SSE, and plain HTTP all work. Activepieces lists the server’s tools as soon as you add it, so you know it works before an agent depends on it.

​
Seeing what an agent did

Actions leave a record. On plans with audit logs, open Platform Admin → Security → Audit Logs and you get the agent, the app, the action, the account it used and whether the run came from chat, the agent page or a flow step. Reads are left out, so the log stays about what changed rather than what was looked at, and the values the agent passed to the action are never stored. A row says whether the action succeeded or failed, so an attempt that errored is still on the record rather than missing. What each tool type leaves behind:
An agent will not change itself in a reply where it already read something. Ask it to update its instructions and it does; ask it to check a sheet and update its instructions in one message and it declines the change, because anything it just read could have told it to. Send the change on its own and it goes through, or make it in the Configure panel.
MCP server tools, code and email are the gaps. MCP tools carry no read-or-write marking, so recording them would fill the log with lookups. If you need a trail for a step, use a piece action or a flow. See Agent action executed for the event itself, and Event Streaming to send it somewhere in real time.
When an agent keeps getting something wrong, ask what it was missing. Usually it needed somewhere to look something up, not a longer prompt.