Skip to main content

Some screenshots show mcp-s.com, which some existing organizations still use. New organizations use withwillow.ai. Copy the URLs shown in your own organization.

Give an agent access to tools

Use this guide to give an existing agent the tools and application authentication needed for its job. You need permission to edit the agent and the relevant MCP server tools available in your organization.

Open Build > Background Agents, select the agent, and open Tools.

Add tools​

  1. Select Add Tools.
  2. On Your MCP servers, search for the server and expand it.
  3. Select the individual tools the agent needs. Tools already assigned are marked Added.
  4. Select Add N tools, where N is the number selected.

The picker also offers Willow connectors, MCP catalog, and Custom API call for other tool sources. To prepare a server first, see Add and configure an MCP server.

Add Tools picker with Your MCP servers selected, a GitHub search, an expanded server, individual tool checkboxes, Added labels, and the Add tools action
On an existing agent, tools already assigned are marked Added. Select additional tools individually.

Choose the integration identity​

Tools are grouped by server. The group shows how many tools are on, and an authentication control such as Auth not set. Open that control to choose how calls to this application authenticate.

GitHub authentication dialog showing Run as the initiator, Use the MCP server default, and Connect my account through Instant OAuth
Identity choices for a GitHub server configured with Instant OAuth.

Run as the initiator​

For a per-user-auth server, choose Run as the initiator when each run should use the triggering person's account. Connect that person's account to the integration first, or complete the connection when the agent requests it.

Slack identifies the triggering person automatically. API triggers and direct MCP/API connections must supply x-user-email or x-user-id. A schedule has no person interacting with it; choose an identity appropriate to the scheduled job rather than assuming initiator authentication is available.

Save the selection. Re-authenticate every run requires a fresh connection on each run when enabled.

Connect your account​

For a fixed account, choose Connect my account for the integration's authentication method. Complete the provider connection, return to the agent's authentication dialog, and Save to point the agent at that connection.

All calls to this integration then use that account. Confirm the account shown in the dialog before saving.

Use a server credential​

Choose Use the MCP server default when the server supplies a shared credential. On a per-user server, this option provides no agent identity and tool calls fail; choose an initiator or connected account instead.

Other credential methods depend on the integration, including API Key, Client Credentials, or Server App. Use the method appropriate to your server. The authentication reference describes these choices.

Verify access​

Confirm the intended tools are enabled and the authentication selection is saved. Then deploy and test the agent with a request that needs one of those tools. Inspect the actual tool result; an assigned tool count does not establish that authentication works.

Rate-limit a tool​

Each row also has a menu (the three dots) at the end. Select Add rate limit and set Allow N calls per minute, hour, or day. The row then shows the limit as a badge, such as 20/hour. The limit counts every call this agent makes to the tool, whoever triggered the run. See Agent Limits for what happens when it is reached.

Require approval for a tool​

On an agent that talks through Slack, the shield icon on a tool row requires the person who triggered the run to approve each call before it runs. Select the icon to turn the requirement on or off. A tool whose MCP server already requires approval shows the icon locked on, because the server enforces it at the source and you cannot clear it here. The icon is not offered on agents that do not use Slack, since the prompt is delivered there.

When the agent calls the tool, it is stopped and Slack posts an Approve / Deny prompt in the thread, right after the agent's reply. The prompt shows the call's arguments. When your organization's content policy redacts content, each value is cut to its first three characters. After the person approves, the agent retries on its own.

An approval covers exactly one call, with the arguments the person saw. A later call to the same tool, or a retry with different arguments, asks again. Without an approval the call is refused.

Keep tools in sync with the MCP server​

An agent reads each tool from the MCP server's tool list in Willow, the list on the server's Tools tab. When that list changes, every agent that uses the server picks up the change on its next run. You do not need to remove and re-add tools, or redeploy the agent.

  • Changed tools. The agent gets the tool's current description and inputs from the server's tool list in Willow, including edits you make there. Its settings on the agent, such as on or off, approval, and rate limits, stay as you set them.
  • Removed tools. The agent stops getting a tool the server's list no longer has. The tool's row stays on the agent's Tools tab with a Removed from MCP badge, so you can remove it or add the tool back on the server.
  • New tools. A tool added to the server is not given to the agent unless the agent follows all of that server's tools.

When the server itself changes a tool's description, Refresh Tools keeps the description already in Willow. Edit the description on the server's Tools tab to pass it to agents.

To have new tools join the agent automatically, open the menu (the three dots) on the MCP server's group in the agent's Tools tab and select Follow all tools. The group shows an All tools badge, and:

  • The agent gets every tool turned on for the server in Willow.
  • When the server is saved or a draft is published, tools added to it join the agent turned on, and tools removed from it leave the agent instead of showing a badge.
  • A tool you turn off on the agent stays off. To keep a tool from the agent, turn it off; tools can't be removed from a server the agent follows.

Select Follow all tools again to stop. The agent keeps the tools it has.

Edit a tool​

Use the edit control beside a tool to open its editor. Save agent-specific changes and deploy them to the runtime where applicable. For schema and mapping authoring, see Edit an MCP tool. The dashboard reference lists agent editor fields.

Generate a tool with AI​

The editor's AI Generate action opens Generate Tool with AI. Describe the required tool behavior, generate it, and review the schema and mapping before saving. Deploy the saved definition where applicable.

Capability requests​

Owners may request additional capabilities when organization settings allow it. Review the request and its explanation, then approve the tool access by adding the appropriate capability and connecting its integration, or dismiss the request. The inventory's Requests tab and request badges identify outstanding requests.

See Organization settings for owner controls and the end-user guide for the requesting workflow.