Connect an MCP client

Applies to: MCP — any client that speaks the Model Context Protocol over HTTP (Claude Code, Cursor, VS Code, and others).

w6w answers as an MCP server at your own host. A model connected to it can find what your account can run, run it, follow a Workflow to the end and check an API’s health — with the same token and the same permissions you use everywhere else.

There are two endpoints. Pick one:

Endpoint	Gives the model
`https://<your-host>/mcp`	Seven discovery and run tools, plus one tool for each Function, Workflow and Endpoint you have published as a tool.
`https://<your-host>/mcp/console`	Everything above, plus five console tools that can read and change anything in your account’s API.

Start with /mcp. Use /mcp/console when you want a model to manage the account itself, not just run what is already in it.

1. Get a token

Studio → Tokens → New token. Treat it like a password: read it from an environment variable and never paste it into a config file.

export W6W_TOKEN=<your-token>

2. Point your client at it

<your-host> is the same origin as the W6W_BASE_URL you use for the SDK and CLI.

Claude Code

Add a .mcp.json file at your project root. Claude Code expands ${VAR} from your shell, so the token never lands in the file:

{
  "mcpServers": {
    "w6w": {
      "type": "http",
      "url": "https://<your-host>/mcp",
      "headers": {
        "Authorization": "Bearer ${W6W_TOKEN}"
      }
    }
  }
}

Or register it from the command line:

claude mcp add --transport http w6w https://<your-host>/mcp \
  --header "Authorization: Bearer $W6W_TOKEN"

Cursor

Add the same shape to .cursor/mcp.json (this project) or ~/.cursor/mcp.json (every project):

{
  "mcpServers": {
    "w6w": {
      "url": "https://<your-host>/mcp",
      "headers": {
        "Authorization": "Bearer ${W6W_TOKEN}"
      }
    }
  }
}

VS Code

Add .vscode/mcp.json to your workspace:

{
  "servers": {
    "w6w": {
      "type": "http",
      "url": "https://<your-host>/mcp",
      "headers": {
        "Authorization": "Bearer ${input:w6w_token}"
      }
    }
  },
  "inputs": [
    {
      "id": "w6w_token",
      "type": "promptString",
      "description": "w6w API token (Studio → Tokens)",
      "password": true
    }
  ]
}

To use the console endpoint, change the URL to https://<your-host>/mcp/console. You can register both under different names.

3. What the model gets

Discovery and run — on `/mcp` and `/mcp/console`

Tool	Does	Hints
`w6w_search_capabilities`	Finds what you can run, by describing the job in words.	read-only
`w6w_list_apps`	Lists apps in the catalogue, and which ones you have connected.	read-only
`w6w_describe`	Shows the exact inputs and outputs of an app, action, Function, Workflow or Endpoint.	read-only
`w6w_connect_app`	Starts connecting an app and hands back a one-time link for you to finish in a browser.	writes
`w6w_invoke`	Runs a Function, Workflow, Endpoint or app action — the same call as `POST /run`.	writes, may be destructive
`w6w_run_status`	Reports how a Workflow run is going.	read-only
`w6w_check_health`	Reports whether an app or connection is working, and how it has been doing.	read-only

A model never sees a credential. To connect an app it calls w6w_connect_app, which returns a link for you to open; the secret stays in your browser and w6w.

A Workflow is not finished when w6w_invoke returns. It answers status: "queued" — started, not done — and the model has to call w6w_run_status to find out how it went.

Console tools — on `/mcp/console` only

Tool	Does	Hints
`w6w_api_search`	Finds the right route for a task.	read-only
`w6w_api_get`	Reads a resource, one or a list.	read-only, idempotent
`w6w_api_write`	Creates or updates a resource.	writes, not destructive
`w6w_api_delete`	Deletes, archives, cancels or otherwise destroys a resource.	destructive
`w6w_guide`	Returns a worked example for a topic.	read-only

Routes that would hand a secret to the model — the vault, tokens, connection credentials and account administration — are left out of the console for every tool. A call that would touch one is refused before it reaches your account. Every tool, with its arguments, is in MCP tools.

Auto-approve

Most clients let you auto-approve tools one by one. Approve only the read-only ones: w6w_search_capabilities, w6w_list_apps, w6w_describe, w6w_run_status, w6w_check_health and, on the console, w6w_api_search, w6w_api_get and w6w_guide. Leave w6w_invoke, w6w_connect_app, w6w_api_write and w6w_api_delete on manual approval — they run real work and change real data in your account.

A worked example

Before you trust a client, check the connection by hand. The endpoint is plain JSON-RPC over HTTP, so curl is enough.

List what the model will see:

curl -sS "$W6W_BASE_URL/mcp" \
  -H "Authorization: Bearer $W6W_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{ "jsonrpc": "2.0", "id": 1, "method": "tools/list" }'

Then call a tool — here, running a Function by its ID:

curl -sS "$W6W_BASE_URL/mcp" \
  -H "Authorization: Bearer $W6W_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{
    "jsonrpc": "2.0",
    "id": 2,
    "method": "tools/call",
    "params": {
      "name": "w6w_invoke",
      "arguments": { "ref": "fn_01H…", "params": { "email": "ada@example.com" } }
    }
  }'

The answer is a JSON-RPC result whose content holds the tool’s reply as text. The ref is one of fn_<id>, wf_<id>, ep_<id>, or an action. Keys and aliases are not refs — use the ID. Pass an action as conn_<id>#<action-key>, or as app:<app-id>#<action-key> to let w6w choose among your connections of that app.

Each server is stateless: there is no session to open or close, and every request carries its own token.

If you would rather make the same call from code than from a model, use client.run — the SDK and CLI do not need MCP.

When it goes wrong

You see	It means	Do this
`401` with `WWW-Authenticate: Bearer error="invalid_token"`	The token is missing, wrong or expired. The body names a `nextAction`.	Mint a new token in Studio and update the header.
`405`	You sent `GET`, `DELETE` or another verb. The endpoint accepts `POST` only.	Use `POST`. A client that probes with `GET` first can ignore this.
A tool result with `isError: true`	The call reached w6w and failed. The text carries a `code` and a `nextAction` hint.	Read the hint. It says whether to retry, fix an argument, or connect the app first.
`unknown_function`, `unknown_workflow`, `unknown_endpoint`, `unknown_connection`	Nothing by that name in your account.	Run `w6w_search_capabilities` or `w6w_describe` again and use a ref it returned.
`invalid_ref`	The `ref` is not one of the five forms.	Use `fn_…`, `wf_…`, `ep_…`, `conn_…#<action>` or `app:<id>#<action>`.
`no_connection`, `connection_pending`, `connection_revoked`, `connection_broken`	The app has no working connection.	Call `w6w_connect_app`, finish the link in your browser, then retry once.
`ambiguous_connection`	You have several connections to the app.	Retry with one of the refs the error lists.
`app_not_entitled`	Your account is not allowed to use that app.	Ask your operator to enable it. Do not retry.
`forbidden_internal_app`	The ref names one of w6w’s built-in apps, which a tool cannot run.	Stop. Do not look for a workaround through another tool.

A failed call carries an invocationId, so you can look it up afterwards — see Invocations.

MCP tools — every tool, its arguments and its results.
Call any API — the same run, over HTTP, the SDK or the CLI.
API health — what w6w_check_health reports.