Setting up

Connect your assistant

One command detects supported assistants, signs you in, and writes the MCP connection and instruction files.

Setup connects both halves of the experience: the MCP capability and the managed instructions that teach your assistant to use it on its own initiative. Afterwards, you can state the outcome instead of coaching every context lookup.

Run the setup

Terminal
# macOS and Linux
curl -fsSL https://contextstream.io/scripts/mcp.sh | bash

# Windows, in PowerShell
irm https://contextstream.io/scripts/mcp.ps1 | iex

It walks you through the whole thing:

  1. Sign in

    Opens a browser. No API key to copy.

  2. Pick a workspace

    Join an existing one or create it. This is the boundary your memory lives in — see Workspaces and projects.

  3. Link this project

    The folder you ran it in becomes a project. Later folders resolve on their own.

  4. Choose your editors

    It detects what you have installed and asks which to connect.

  5. Finish the connection

    The setup creates each selected editor's connection and instruction files for you. Where supported, it also enables automatic context guidance.

  6. Index the project

    Starts in the background. You can work while it runs.

Restart your editor afterwards. That's it.

What changes after setup

Your assistant can now orient to the active project, retrieve context relevant to the current request, search the appropriate evidence, and preserve durable work when you ask. The generated instructions also tell it to stop when scope is ambiguous and to treat matching lessons as working guardrails.

ContextStream does not widen the assistant's execution permissions. You still control the goal and consequential actions; the connection removes repeated project explanation, not human authority. See how the agent partnership works for the full loop.

Check it worked

Ask your assistant something that requires project context from the connection:

Code
what project am I in?
what did we decide about anything so far?

The first should name your project. The second will be empty on day one — that's the correct answer, and it proves the connection.

Your editor's exact details

The setup wizard currently manages Claude Code, Cursor, GitHub Copilot in VS Code, Windsurf, Cline, Kilo, Roo, Codex CLI, Aider, Antigravity, and OpenCode. For each one you select, the wizard creates every supported connection and instruction file automatically. On editors with lifecycle support, it also installs the automation that surfaces relevant context and guardrails while you work. Aider supports instruction rules only. Claude Desktop is supported through manual configuration, but is not one of the wizard-managed targets.

Pick yours to see its exact files and connection shape:

Claude Code

Hosted connection

Anthropic's terminal and IDE coding agent.

Connection file
mcp.json~/.claude/mcp.json globally, or .mcp.json in a project
Instructions file
CLAUDE.mdProject root, or ~/.claude/CLAUDE.md globally
mcp.json
{
  "mcpServers": {
    "contextstream": {
      "type": "http",
      "url": "https://mcp.contextstream.io/mcp?default_context_mode=fast",
      "headers": {
        "X-ContextStream-API-Key": "YOUR_API_KEY"
      }
    }
  }
}

Replace YOUR_API_KEY with a key from your dashboard, keep the file out of version control, and restart the client. The setup command writes the fuller managed configuration for you.

Connecting more editors later

Run the setup again — it's safe to re-run and won't duplicate anything. Or use the maintenance commands in Keeping it current.

If it doesn't work

The assistant doesn't seem connected. Restart the editor. Most read MCP configuration only at startup.

It connected but knows nothing about the project. Indexing runs in the background and takes a few minutes on a large repository.

The script can't find my editor. It looks in standard install locations. Set it up by hand — see Set it up by hand.

Next