Skip to main content
Choose this setup when you want your agent to prefer Context for public-web searches and page reads without naming it each time. Highlights return relevant passages; full Markdown is available when you need the whole page. Other MCP integrations stay enabled unless you approve changing them. Want to keep your current defaults? Connecting your agent is enough; you can skip this page.

​
Ask your agent to set it up

Paste this into the connected client. It keeps the same project or personal scope; ask explicitly if you want to change that scope.
Setup prompt
Setup uses a few Context credits and may require browser sign-in or a restart. You can also follow the steps below yourself.

​
1. Verify your connection

If you have not connected yet, connect your agent first. Reuse its existing connection to https://mcp.context.dev/mcp; a Mintlify documentation-search server does not provide Context’s web tools. Discover the connected tools and inspect their schemas rather than assuming names. Reuse a successful connection test from this session. Otherwise, ask the agent to scrape https://example.com through Context and confirm success before disabling any working web tools. If it fails, fix the connection first.

​
2. Configure your client’s defaults

​
Claude Code

Merge these built-in tools into permissions.deny in ~/.claude/settings.json, preserving existing entries:
For personal project-only settings, use .claude/settings.local.json. Honor CLAUDE_CONFIG_DIR if customized. See Claude Code settings.

​
Codex

Set this top-level key in ~/.codex/config.toml, before any [table] header. Replace an existing value instead of duplicating it:
Honor a custom CODEX_HOME. For project-only setup, use .codex/config.toml in a trusted project. Check profiles, project overrides, and launch flags such as --search that can override this setting. See the Codex configuration reference.

​
Other MCP tools and clients

Ask before disabling or removing another MCP integration. Show the user the affected server, exact tools, and capabilities they would lose. Get explicit approval for those specific changes; leave unapproved tools enabled. After approval, use exact discovered tool names in Claude Code’s deny list, or exact server-side names in that server’s Codex disabled_tools list. Setting enabled = false requires approval to disable the entire server, including its other tools. Inspect plugin-provided configuration before editing; do not add a duplicate server. For other clients, use their supported controls for built-in web tools. If controls are unavailable, the instructions below establish a preference, not guaranteed routing. Keep repository tools, private-data connectors, shell access, and browser automation available. Verify each client separately.

​
3. Save routing instructions

Merge the following into the instruction file the client loads:
  • Claude Code: ~/.claude/CLAUDE.md.
  • Codex: ~/.codex/AGENTS.md. An existing AGENTS.override.md takes precedence.
  • Project-only: the project’s instruction file.
Honor custom configuration directories, preserve unrelated instructions, and check for more-specific conflicting rules.
Persistent instructions

​
4. Verify and keep undo steps

Start a fresh session and run both checks without naming Context:
Search check
Page-reading check
Check actual tool activity: both must use Context successfully, with Highlights on the page read. Limit verification to the initial connection scrape and these two checks; do not retry automatically. A saved configuration or a failed call is not proof. If a restart or sign-in is still needed, report verification as pending. To undo, restore the previous settings and remove only the routing instructions and tool filters added during setup. Preserve pre-existing deny entries and other settings. Keep the Context connection if you still want it available on request, then restart the client.