Connect Cursor, Claude, Codex, VS Code, ChatGPT, and any MCP-compatible agent to your CitePatch workspace — projects, visibility metrics, opportunities, Brand Truth, and patches. One secure endpoint. One-click OAuth. No API keys to copy.
https://mcp.citepatch.com/mcpModel Context Protocol
The Model Context Protocol (MCP) is an open standard that lets AI assistants talk to tools and data through one consistent interface. CitePatch exposes your AI Search Patch Loop as an MCP server, so your agent can read evidence, draft patches, and submit them — without copy-paste.
It runs on Streamable HTTP at a single endpoint and authenticates with OAuth 2.1 — your client opens CitePatch consent, you pick the organization, and you approve exactly which scopes to grant. No API keys to create or paste. Write actions are confirmation-gated.
Your agents
OAuth 2.1 · PKCE
MCPmcp.citepatch.comCitePatch
You approve scopes at sign-in. Grant only what you need — anything outside them is rejected.
read scopesSee your workspace the way CitePatch sees it — without leaving the agent.
patches:writeDraft content changes in the agent, then send them into CitePatch.
patches:applyWhen the org allows it, open a reviewable PR or draft from the agent.
Built-in prompts (brief-me, close-gap) walk the agent through briefing and closing a gap step by step.
Point your MCP client at https://mcp.citepatch.com/mcp — a CLI command in Claude Code, a JSON line in Cursor, or an HTTP entry in Codex. Configs below.
Your client opens CitePatch OAuth in the browser. Sign in, choose the organization, review scopes (read, write, apply), and Approve.
The client stores a short-lived token with the org bound in. Ask your agent about visibility gaps, evidence, or patch drafts — no headers to configure.
Point any MCP client at https://mcp.citepatch.com/mcp (Streamable HTTP). The first call triggers CitePatch OAuth — pick an org, grant scopes, done. No API keys, no custom headers. For inspection only, use https://mcp.citepatch.com/mcp/readonly.
Drop citepatch into your global MCP list, or commit a project file so the whole team gets the same connector.
{
"mcpServers": {
"citepatch": {
"url": "https://mcp.citepatch.com/mcp"
}
}
}Desktop builds that only speak stdio need a local bridge — mcp-remote forwards HTTP and runs OAuth for you.
{
"mcpServers": {
"citepatch": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://mcp.citepatch.com/mcp"]
}
}
}Any stdio-only client can reuse this npx mcp-remote wrapper against the same URL.
Register the remote server once from the shell. Pass --scope user if you want it on every project.
claude mcp add --transport http citepatch https://mcp.citepatch.com/mcpWorks for user-level MCP and workspace configs — including Copilot Chat when MCP is enabled.
{
"servers": {
"citepatch": {
"type": "http",
"url": "https://mcp.citepatch.com/mcp"
}
}
}Schema quirk: top-level key is servers, and each entry needs type: "http" (not the mcpServers shape Cursor uses).
Add CitePatch as a connector in ChatGPT settings — URL plus OAuth, nothing to paste as a secret.
https://mcp.citepatch.com/mcpDeclare an HTTP MCP server in your Codex config, then reload so discovery can run.
[mcp_servers.citepatch]
url = "https://mcp.citepatch.com/mcp"OAuth over MCP is still uneven across Codex builds — if discovery stalls, Cursor or Claude Code are the safer path.
Once connected, talk to your agent in natural language. It picks the right tools.
“Use citepatch_get_workspace then brief me on our top visibility gaps for the main project.”
“List opportunities and show evidence for the highest-impact owned gap.”
“Get the patch brief for opportunity X and draft a minimal content change grounded in Brand Truth.”
“Show experiment ledger entries and whether any patch correlated with a visibility change — do not claim causality.”
MCP calls send your AI client exactly what each tool returns — no raw embeddings dump, no internal metadata. Write tools need their own scope and leave an audit trail. We do not use your workspace data to train third-party models.
Add the endpoint, approve access in one click, and let your agent work the Patch Loop with real evidence.