Skip to main content

Connect Claude Tag

On the Claude Tag platform (claude-tag) the agent doesn't run on a Willow-managed runtime. Claude itself is the runtime: you add the agent to a Claude bundle as a connected app, and it runs whenever someone mentions @Claude in a Slack channel the bundle applies to.

Willow's job on this platform is to expose the agent's tools and to authenticate Claude to the gateway. Nothing is pushed to Anthropic, so there is no Deploy step — the connection is configured entirely in Claude, using values Willow shows you.

Two ways to connect​

MethodCredential type in ClaudeWhen to use it
MCP connectorMCP ConnectorThe default. Claude connects to the agent as an MCP server and discovers its enabled tools itself, so you don't maintain an instructions snippet.
REST APIBearerFor bundles that can't hold an MCP connector. Each tool is its own POST endpoint, and you paste an instructions snippet telling Claude which endpoint to call.

Both surfaces expose the same tools — the agent's enabled tool capabilities, scoped and logged by the gateway exactly like any other MCP traffic. Skills, rules, and webhook triggers are not available on either.

Both are set up from the agent's Settings → Platform Configuration → Connect to @Claude card, which has a tab per method. Each tab shows the endpoint, the credential fields to enter in Claude, and a How to connect walkthrough.

Authentication​

Whichever method you pick, Claude authenticates with the agent's own gateway credentials:

Authorization: Bearer <access_key>:<secret>

The access key is on the agent's Settings tab. The secret is shown once, when the agent is created or when you rotate it — copy it before leaving the page. Rotating the secret invalidates the value stored in Claude, so update the connected app after every rotation.

Connect as an MCP connector​

  1. In Willow, open the agent's Settings → Platform Configuration, select the MCP Connector tab, and copy the MCP Server URL. It ends in /agent/<agent-slug> on your gateway host, for example https://acme.mcp-s.com/mcp/agent/support-bot.
  2. In Claude, open your bundle's Credentials → Connect an app.
  3. Set Credential type to MCP Connector, then choose the Provider matching your gateway host and confirm the MCP server URL.
  4. Under Authentication, switch from MCP Connector to Custom authentication. This connects with fixed headers instead of an interactive sign-in, which is what a background agent needs.
  5. Add one header, then select Connect:
FieldValue
Header nameAuthorization
Header prefixBearer
Header secret<access_key>:<secret>
Header values are write-only

Claude never displays a header secret again after saving. If you lose it, rotate the agent secret in Willow and re-enter the header.

Mention @Claude with the bundle applied and it lists the agent's enabled tools from the connector, calling them through the gateway.

Connect over the REST API​

  1. In Willow, select the REST API tab on the same card. It shows the base endpoint, which ends in /agent-api/<agent-slug>/{tool}, and the Allowed website (the gateway host).
  2. In Claude, open Credentials → Connect an app and pick Bearer as the credential type.
  3. Enter the allowed website and the Authorization header shown in Willow.
  4. Open the bundle's Instructions tab and paste the snippet from How to connect. It lists one POST endpoint per enabled tool, with each tool's description, so Claude knows what to call and when.

Each tool is reachable at that base URL plus /<integration-slug>__<tool-slug>, called with POST and a JSON body matching the tool's input schema; the response is the tool result. The API reference button on the card lists every endpoint with a ready-to-run curl example and its body schema.

What this platform doesn't support​

SurfaceStatus
ToolsSupported, over both the MCP connector and the REST API.
SkillsNot supported. Both surfaces expose tools only.
TriggersNot supported. The agent runs when someone mentions @Claude; it has no API, scheduled, or webhook triggers.
SessionsNot recorded. Claude owns the conversation; individual tool calls still appear in monitoring.
DeployNot applicable. Nothing is synced to Anthropic.
On this page