Setting up

Connect ContextStream Custom MCP

This is a custom MCP connector. Paste the hosted URL anywhere custom MCP connectors are supported, then sign in with OAuth.

This is a custom MCP connector. Any product that lets you add a custom MCP connector — sometimes labeled custom connector, remote MCP, or custom server — can use the same hosted ContextStream URL. Claude.ai and Grok are worked examples, not the only clients.

The connected assistant then reaches the same project record your editors use: indexed source, decisions, lessons, plans, runbooks, past sessions, and the rest of the durable context.

You paste one URL. The client discovers authorization on its own. There is no API key to copy, no client ID, and no local config file.

Code
https://mcp.contextstream.io/mcp

“what did we decide about the billing migration?”

The captured decision from last week's editor session, not a fresh guess.

“what's still open on token refresh?”

The plan, tasks, and the snapshot where work stopped.

“write the postmortem from what actually happened”

A draft grounded in sessions, lessons, and the runbook — not reconstructed from memory.

“remember we're staying on one shared refresh request”

Saved as durable project context your editors will see tomorrow.

Anywhere custom connectors are supported

Look for the setting that adds a custom or remote MCP server, paste the URL above, and complete the ContextStream sign-in the client presents. Leave client ID, client secret, and API key fields empty unless you registered a private OAuth client.

Claude.ai and Grok have the steps below. Another chat app, research tool, or agent host that supports custom MCP connectors takes the same URL and the same OAuth flow. You do not need a per-product ContextStream integration.

If the client only accepts a local command, a stdio binary, or an API key in a header, that is a different path — use Connect your assistant or Set it up by hand.

Claude.ai

On a Pro or Max plan:

  1. Open connectors

    In Claude.ai, go to Customize → Connectors. Some accounts label this Settings → Connectors.

  2. Add a custom connector

    Click +, then Add custom connector.

  3. Paste only the URL

    Use https://mcp.contextstream.io/mcp. Leave Advanced settings empty. Do not add a client ID, client secret, or API key.

  4. Sign in

    Add the connector. Complete the ContextStream sign-in window that opens.

  5. Enable it in the chat

    In a conversation, open + → Connectors and turn ContextStream on for that thread.

On Team and Enterprise, an owner adds the URL first under Organization settings → Connectors → Add → Custom → Web. Each member then chooses Connect from Customize → Connectors and signs in with their own ContextStream account. Claude can only see the workspaces that person can already access.

Grok

  1. Open connectors

    Go to grok.com/connectors.

  2. Add a custom connector

    Click New Connector, then Custom.

  3. Paste the same URL

    Use https://mcp.contextstream.io/mcp and complete the ContextStream sign-in Grok presents.

On Grok Business and Enterprise, a team admin provisions the connector before members can use it.

What you get that a normal chat does not

The connection is the production Rust MCP — the same tool surface your supported editors use. After you sign in, the connected client can:

  • Search the indexed repository instead of guessing file names.
  • Read decisions, lessons, preferences, and facts that were captured in another assistant.
  • Resume a plan, a task, or a session snapshot from last week.
  • Draft from a runbook, a diagram, or an incident instead of a blank page.
  • Save a new decision, lesson, or note that your editor will see later.

That is the point of a custom connector. You can write the announcement, plan the quarter, or review an incident in a browser and still be looking at the project, not at a generic model.

A Monday-morning example: last Thursday in Cursor you recorded why token refresh is coalesced, left the clock-skew test failing, and saved a snapshot. Today you open Claude.ai, enable the connector, and ask it to draft the billing postmortem and list what is still open. It should cite that decision, the failed verification, and the open task. A chat without the connector invents a plausible postmortem; this one starts from the record.

You still set the outcome and authorize consequential writes. The assistant should stop when scope is ambiguous or the next action would change something you did not ask it to change.

What it cannot do

These clients have no view of your filesystem. They cannot edit the repository, run tests, or index a folder that has never been synced. Use a coding editor connected through the setup wizard for changes and for the first index of a project.

They also have no managed instruction file. Ask for project context when you need it — "check our decisions first" — until the assistant learns the habit. See What to say once it's connected.

How sign-in works

Paste the URL and nothing else. The client treats that address as a remote MCP server, discovers that it is OAuth-protected, registers itself if it needs to, and opens ContextStream's sign-in page. After you approve access, later requests carry a token. You do not create an API key for this path.

If it doesn't work

The connector asks for a client ID or secret. Leave those fields empty and save again. Filling them skips dynamic registration and usually breaks sign-in.

Sign-in never completes. Allow the popup, then retry Connect. If you use a Team or Enterprise Claude workspace, confirm an owner already added this exact URL.

It is connected but answers generically. Enable the connector on that conversation, then ask it to check project decisions or name the current project. With no instruction file, the first prompt has to request context.

It knows nothing about the code. Indexing happens from a machine that ran setup, not from the web chat. Memory captured in other connected assistants is still available; source search stays empty until a checkout has been indexed.

Next