Connect your agent

Configure Toolstile as a remote MCP server in your agent application, then sign in through the browser and choose a workspace. That authorization controls which workspace the agent uses. After approval, grant the agent access to the MCP connections it needs; sign-in alone does not grant it tools.

Before starting, add a public HTTPS MCP connection in the workspace you want to use. Sign-in to that upstream service is separate from the agent's Toolstile authorization.

The gateway URL is:

https://toolstile.com/mcp

Use this gateway URL in your client, not the upstream server's URL or an /app/… dashboard address. Connect an agent in Toolstile also gives you the configuration for your chosen client.

Claude Code

Add the remote HTTP server from a terminal:

claude mcp add --transport http toolstile 'https://toolstile.com/mcp'

The default scope is local to the current project. Choose a user-level scope only if you want that configuration available across projects; Claude Code's MCP reference explains the scope options.

Open Claude Code and run /mcp. Select Toolstile and follow its authentication prompt. In the browser, choose the Toolstile workspace and review the proposed agent/device names before approving.

Return to Toolstile to grant the new agent access, then reconnect or refresh Claude Code's catalog. Adding the server does not authorize other Claude Code entries on the same computer.

Codex CLI

Add the URL and start browser authorization:

codex mcp add toolstile --url 'https://toolstile.com/mcp'
codex mcp login toolstile

Choose the workspace in Toolstile's approval flow. The CLI stores the server entry in ~/.codex/config.toml; avoid adding a second table for the same entry. Project-specific configuration belongs in a trusted project's .codex/config.toml.

After granting access in Toolstile, use codex mcp list to inspect the configured entry and /mcp inside Codex to inspect active servers. These are local Codex CLI settings; they do not configure a hosted ChatGPT conversation. See OpenAI's MCP guide.

Cursor desktop editor

Add a project-specific entry to .cursor/mcp.json, or use ~/.cursor/mcp.json for a personal configuration across projects:

{
  "mcpServers": {
    "toolstile": {
      "url": "https://toolstile.com/mcp"
    }
  }
}

Merge this entry into your existing mcpServers object. Restart Cursor, open Customize → MCPs, and follow Toolstile's sign-in prompt. Choose the workspace and review the suggested agent name in the browser.

Grant access in Toolstile before asking Cursor to use a tool. A project entry overrides a global entry with the same name. These instructions cover the desktop editor; other Cursor surfaces have their own configuration. See Cursor's MCP reference.

Claude Desktop

Use Claude's remote custom-connector flow. Open Customize → Connectors, choose Add custom connector, give it a recognizable name, and enter the gateway URL. Your Claude account or organization must permit custom connectors; availability and controls depend on its plan and role.

Choose a sign-in option rather than No sign-in. If Claude asks how it should identify its OAuth client, Toolstile accepts both Use Claude's published identity and Register automatically, so you do not need a client ID or secret. A Toolstile agent token is not an OAuth client secret.

Approve the intended Toolstile workspace, then grant the new agent access in Toolstile. In a Claude conversation, open + → Connectors and enable the entry before requesting a tool.

A remote connector belongs to your Claude account and sends requests from Anthropic's infrastructure. It is different from a local MCP process configured in claude_desktop_config.json. A Toolstile device label groups agent connections; it does not prove requests originate from that computer. A label such as “Claude account” can make that distinction clearer.

See Claude's custom-connector instructions. Dialog labels can differ between accounts.

Grant access and check the result

Find the new agent in Devices & agents in the workspace you approved. In Access, choose the MCP entry and the capabilities the agent needs, then save. In a team workspace, an administrator grants each agent access individually.

For example, allow a coding agent to look up an issue while leaving write tools unavailable. Full access also includes future capabilities from that server; explicit blocks still win. Review permissions before choosing it.

Start with Toolstile's permitted-catalog check. It does not execute a tool. To check the complete client path, explicitly ask the client for a read-only lookup using non-sensitive input. Review Activity in the intended workspace for the expected agent and connection. Tool names shown through the gateway can differ from the upstream names.

Personal and team contexts

Each authorization belongs to one workspace. Switching the Toolstile dashboard does not retarget an existing client authorization.

Keep separately labeled configurations for personal and team work. A label alone does not establish separate credentials: check that the client completes a distinct approval and that the resulting agent appears in the intended workspace. If it reuses an existing sign-in, resolve that before relying on the second context. Client-specific connector limits can also apply.

Token setup and revocation

For a script or client that accepts a bearer token, Connect an agent → Create a token instead issues a separate credential in the selected workspace. Choose its expiry, copy the token (it is shown once), and grant that agent access. Keep the value in a private secret source, outside project files and version control.

Token configuration depends on the client. Do not paste a generic header block into a client that only exposes an OAuth connector, or treat an agent token as a provider API key. The detailed client references above describe their authentication options.

Choose Remove agent in Toolstile to stop that agent's workspace access. Use Remove device only when you intend to stop every agent grouped under it, in every workspace. Revocation cannot reverse actions already completed upstream.

If sign-in succeeds but tools are missing, check the workspace, saved grant, upstream authentication and selected account. Refresh the client's catalog after access changes. Continue with troubleshooting for quota stops or an unavailable account.