Why does the local Tares MCP endpoint not respond after tares up, and how do I start it?

Tares does not serve /mcp from the daemon alone on your machine, so you start the MCP server as a second process with tares mcp. Run it with the streamable-http transport, port 8788 and the —taresd URL of your daemon, and http://localhost:8788/mcp starts answering. On your own server or on a Tares Cloud instance the MCP server is already running, so there is nothing to start.

How it works

Running tares up starts the Tares daemon, which listens on http://localhost:8787. The daemon does not serve the MCP endpoint. The MCP server is a separate process that listens on its own port, 8788, and forwards what your agent asks to the daemon. That is why a request to http://localhost:8788/mcp gets no response until you start it.

Where Tares runs decides what you do:

• Your machine (tares up): the endpoint is http://localhost:8788/mcp. Start the MCP server yourself, as shown below.
• Your own server (compose deployment): the endpoint is your domain followed by /mcp over HTTPS. The MCP server is already running.
• Tares Cloud: each instance has its own hostname, shown in your console under Connect. The MCP server is already running.

The --taresd flag tells the MCP server which daemon to talk to. For a local setup that is the daemon’s address on port 8787. The streamable-http transport makes the MCP server a network endpoint that clients connect to over HTTP. It is the default path for connecting agents and the only one for remote instances.

The full list of clients and tools is on the Connecting agents (MCP) page. If you use Tares for incident work, the AI SRE guide shows what an agent does once it is connected.

Set it up

1. Start the daemon in your first terminal.

   tares up

2. In a second terminal, start the MCP server. Keep this terminal open, because the endpoint stays up only while the process runs.

   tares mcp --transport streamable-http --port 8788 --taresd http://localhost:8787

3. Connect your client to http://localhost:8788/mcp. For Claude Code, run this in a third terminal:

   claude mcp add --transport http tares http://localhost:8788/mcp

   Cursor takes an MCP server entry with a url key in mcp.json. Codex uses codex mcp add tares --url http://localhost:8788/mcp. The Connecting agents (MCP) page has the exact snippet for each client, including VS Code, Windsurf, Cline, Zed and Gemini CLI.

4. Verify the connection. Ask your client:

   Use tares: what are you ingesting right now?

   The client calls catalog_list and answers with your sources, triggers and projects. In the console, Reads lists each read as it happens, tagged client = mcp.

If you installed the Tares plugin for Claude Code, it asks for the MCP endpoint during install and defaults to http://127.0.0.1:8788/mcp. The same rule applies: the MCP server from step 2 must be running first.

On a secured instance, started with tares up --auth, the MCP surface requires a credential. The client sends an Authorization Bearer header holding an API key, and a key with the read scope is enough for reading. On a plain tares up instance no credential is needed.

What you get

• A working local endpoint at http://localhost:8788/mcp that any MCP client can use.
• Your agent can read the correlated timeline across your sources, list sources, triggers and projects, and use the other Tares tools your key’s scope allows.
• A clear picture of which setup needs a second process: only a local tares up. A server deployment or a Tares Cloud instance already serves /mcp.
• Reads from your agent show up in the console under Reads, so you can confirm the connection works.

When something else fits

• If your agent runs on the same machine as the daemon and you would rather not run an HTTP server, use the stdio transport. The client spawns tares-mcp as a subprocess that proxies to the daemon at the TARESD_URL environment variable.
• If you use a server deployment or Tares Cloud, skip the second process and copy the endpoint from your deployment or the console’s Connect page.
• If you only want to explore your data without an external client, open Ask in the console. It uses the same read tools inside Tares.

Related