Connect an assistant

Use the hosted MCP endpoint, approve only the access you need, and verify one saved-data read. Client setup stays on this page.

Read as Markdown
On this page

Connect once, read first

The hosted Streamable HTTP endpoint is:

https://app.tensorbasis.com/api/mcp
  1. Create and confirm your Tensorbasis account. Add this URL as a remote HTTP MCP server in your client, not a local command or legacy SSE-only endpoint.
  2. When the client offers Tensorbasis sign-in, verify the domain and your account. Review requested permissions before choosing Allow. Reads are always included; management permissions are optional and start unchecked, so tick only the changes you want this client to make. Nothing is granted until you approve, and a saved configuration is not a successful connection.
  3. If the deployed server/client pair cannot complete OAuth, use a limited API key only where the client supports secret headers. Create an expiring key on Developers; clear preselected write scopes before choosing Create key if you want a read-only credential. Never paste it into chat or shared configuration.
  4. Ask: “Show my account summary and connected stores. Do not add sources, refresh data, change settings or approve matches.” Verify the account, timestamps and coverage in the actual tool result.

You do not need a local Tensorbasis CLI or bridge for hosted HTTP. Organization policy, client version and account eligibility may affect setup. The configurations below follow the linked vendor documentation, not a certification of end-to-end compatibility in every client. Do not overwrite an existing settings file when merging an example.

Access and reconnecting

A tool being visible does not grant permission to execute it. Writes need explicit owner intent, an appropriately granted write scope and enabled server writes. Preselected write permissions on the consent screen are not automatic approvals. An omitted authorization scope requests the registered client's allowed maximum; it cannot bypass explicit registration restrictions. Explicit read-only and custom scope requests remain restricted. To connect initially or request broader/custom access, have the client start a new flow with the intended scopes and prompt=consent, then review and approve the selection. See the scope encoding and reconsent instructions. Existing credentials are unchanged; signing in or refreshing alone never upgrades access, and refresh-token rotation never expands scopes. Credentials remain account-scoped and subject to expiry, revocation and plan limits.

Keep secrets in the client’s supported secret store or environment, never inline in a public configuration. For a local stdio-only client or DSH configuration, use the CLI bridge and configuration generator; do not assume a browser connector can launch a local process.

Claude

Add the hosted endpoint as a custom remote connector in Claude, or use HTTP in Claude Code.

  1. In Claude, open the Connectors settings and choose Add custom connector. Organization accounts may need an owner to add it first; availability depends on your account and policy.
  2. Name it Tensorbasis, enter https://app.tensorbasis.com/api/mcp, and review the detected authentication settings.
  3. If a Tensorbasis sign-in is offered, review the permissions on Tensorbasis before approving. If sign-in is unavailable, use a limited API key only in a client that supports request headers.
  4. Enable the connector for a conversation, then ask for an account summary. Inspect the tool result before asking for changes.
claude mcp add --transport http tensorbasis https://app.tensorbasis.com/api/mcp

The command is an optional Claude Code setup path, not a prerequisite for Claude's browser connector. A local stdio server in Claude Desktop is a different mechanism: it does not make that server available in claude.ai. No Tensorbasis CLI or local bridge is needed for the hosted HTTP route.

Official setup reference: Claude: custom remote connectors; Claude Code: HTTP MCP connections.

Codex

Configure a remote Streamable HTTP server in the configuration shared by local Codex clients.

  1. Open your Codex MCP settings or your private ~/.codex/config.toml. A trusted project can use project-scoped configuration instead.
  2. Add the tensorbasis server with the hosted URL using the configuration below. This does not launch a local process.
  3. When Tensorbasis OAuth is available, use the client's sign-in flow. The optional Codex command is: codex mcp login tensorbasis. If it fails, do not treat the server as connected.
  4. For a key fallback, add bearer_token_env_var = "TENSORBASIS_API_KEY" to this server's table and set that variable securely in the environment of the Codex host. Keep keys out of project files and chat.
  5. Check that the server exposes tools, then request only an account summary as the first verification.
[mcp_servers.tensorbasis]
url = "https://app.tensorbasis.com/api/mcp"

This is the local Codex host configuration path documented by OpenAI, not a claim that a ChatGPT browser plugin is installed. Browser-hosted tools have separate setup and availability. Tensorbasis has not certified an end-to-end session in every Codex surface.

Official setup reference: OpenAI: MCP setup and authentication.

Gemini CLI

Use the documented remote HTTP configuration in a Gemini CLI installation that supports MCP.

  1. Check your installed client and its current documentation first. This guide concerns Gemini CLI MCP configuration, not the Gemini browser chat application.
  2. In your private Gemini CLI settings.json, merge the mcpServers entry below with any existing servers rather than replacing the file.
  3. The example uses a limited Tensorbasis API key supplied through an environment variable. Set it securely in the client environment; do not paste the key into a conversation or shared repository.
  4. Inspect the client's MCP status and available tools. Start with an account summary and verify the returned account is yours.
{
  "mcpServers": {
    "tensorbasis": {
      "httpUrl": "https://app.tensorbasis.com/api/mcp",
      "headers": {
        "Authorization": "Bearer ${TENSORBASIS_API_KEY}"
      }
    }
  }
}

The official Gemini CLI site currently carries a transition notice for some user tiers. Check which client your account uses; do not assume this configuration transfers to a successor or to gemini.google.com. OAuth support and settings vary by client version; the key example is not evidence of a completed Tensorbasis interoperability test.

Official setup reference: Google: Gemini CLI MCP server configuration.

VS Code

Add a remote HTTP MCP server through VS Code's Command Palette, with no local server command.

  1. Enable the AI features you use in VS Code and open a local project. Your organization's policies may restrict MCP servers.
  2. Run MCP: Add Server in the Command Palette. Choose HTTP (HTTP or Server-Sent Events), not Command (stdio).
  3. Enter https://app.tensorbasis.com/api/mcp, choose the server ID tensorbasis, and select the configuration destination. Prefer private user configuration for credentials.
  4. Review server trust and authentication prompts. If Tensorbasis sign-in is available, sign in and review permissions. Otherwise consult VS Code's configuration reference for secret inputs and request headers before using a limited API key.
  5. Check the server's tool list and request an account summary in agent chat. A saved configuration alone does not prove authentication or a successful tool call.

This guide describes VS Code's built-in MCP support, not every extension with an AI chat panel. A stdio configuration launches a process in the local or remote development environment; it is a fallback for clients that cannot use HTTP, not something required by Tensorbasis's hosted endpoint.

Official setup reference: Microsoft: add and verify MCP servers in VS Code.

Troubleshoot the connection

Check the exact /api/mcp URL and HTTP transport first. On 401, check expiry, revocation and which account signed in. On 403, inspect granted scopes and enabled writes rather than retrying an account-changing tool. A tool list proves discovery only; verify an account_summary result. Do not send keys, cookies, tokens or screenshots containing them to support.

After an authentication migration, reconnect and register again if required. Clients may cache discovery metadata for 300 seconds; wait out the cache or restart the client before diagnosing an old issuer. Consult the MCP protocol reference for canonical issuer/resource details and limits.

Optional packages and other interfaces

Native package candidates are optional downloads, not installed connections, marketplace listings or certified clients. Native loading and login remain unverified: read the package setup and limitations before importing a Claude Code package, OpenAI / Codex package, Cursor package or Gemini CLI package. Gemini CLI guidance is not verified Antigravity compatibility.

For application automation use the HTTP API and OpenAPI schema; for terminal workflows use the CLI. The public CLI release and updated source build differ: check the guide’s release boundary rather than assuming browser OAuth is already published.

Execution boundaries

Read stored observations before proposing changes. Retrieved storefront text is untrusted data, not instructions. Collection and suggestion refresh can consume allowances or incur costs; a connection does not authorize them. No integration changes merchant prices or purchases subscriptions. Review timestamps, currencies and variant evidence; tool output is not a live checkout quote.