MCP Connection

Connect Cursor, Claude, or any MCP client to EgoX to manage your tools & knowledge.

EgoX runs a hosted MCP server at https://mcp.egox.io. Point any Model Context Protocol client — Cursor, Claude, or your own agent — at that URL and it can build, update, and review your project's tools and knowledge in natural language, straight from your editor.

There's nothing to install or run. It's a remote endpoint you connect to over Streamable HTTP / SSE.

Cursor / Claude / your agent ──SSE (Streamable HTTP)──► https://mcp.egox.io

Connect

Add EgoX to your MCP client's server list. There are two ways to authenticate.

Point your client at the bare URL — no token needed. On first connect it opens a browser window where you sign in, pick a project, and approve the connection.

{
  "mcpServers": {
    "egox": { "url": "https://mcp.egox.io" }
  }
}

That's it — the client handles the sign-in handshake for you. The connection it creates is listed on your project's MCP Tokens page in the Console, where you can revoke it anytime.

Prefer a fixed credential (e.g. for a headless agent or CI)? Generate an MCP token in the Console — open your project → MCP Tokens → create a token with a name and expiry, then copy it (it's shown only once).

Pass it on the connection URL:

{
  "mcpServers": {
    "egox": { "url": "https://mcp.egox.io?accessToken=egox_mcp_xxxxxxxxxxxxxxxx" }
  }
}

Or, if your client supports custom headers, send it as Authorization: Bearer egox_mcp_… instead of putting it in the URL.

Restart your client and the egox_* tools appear in chat.

Which clients work

Any MCP client that speaks Streamable HTTP (recommended) or the legacy HTTP+SSE transport. The stream is resumable, so a dropped connection reconnects without losing events. Cursor and Claude are supported out of the box.

What the agent can do

Once connected, EgoX exposes these tools to your agent:

Tool	What it does
`egox_list_tools`	List all tools in the project
`egox_get_tool`	Get a single tool by `toolName`
`egox_upsert_tool`	Create or update a tool by `toolName`
`egox_delete_tool`	Delete a tool by `toolName`
`egox_list_rag_documents`	List knowledge-base documents
`egox_get_rag_document`	Get a knowledge doc by `title`
`egox_upsert_rag_document`	Create or replace a knowledge doc by `title`
`egox_delete_rag_document`	Delete a knowledge doc by `title`

Anything the agent creates is tagged source = 'mcp', so the Console surfaces it as a draft for you to review. To overwrite content a human curated, the agent must explicitly pass force: true — a guard against accidental rewrites.

Tool-definition fields the agent can set

When the agent creates or updates a tool, these are the knobs it can control:

Field	Purpose
`inputSchema` / `exampleInput` / `exampleOutput`	What the LLM sends + reference examples
`headers`	Static headers EgoX always sends (e.g. an upstream API key)
`forwardHeaders`	Per-request allowlist of inbound `/ask` headers to forward (e.g. `["x-locale"]`). Sensitive headers (Cookie, Host, Authorization, X-Api-Key, …) are blocked
`authType`	`FORWARD` passes the caller's `Authorization: Bearer …`; `NONE` sends no auth
`gqlQuery`	GraphQL operation string (use `$variables`, never hardcoded literals)

What you can ask the agent

Once connected, prompt it like a teammate:

"Look at the OrdersResolver in this subgraph and create an EgoX tool listMyOrders that calls the corresponding GraphQL query. Use $userId as a variable, never hardcoded."

"Update the getOrder tool to add an optional includeItems boolean argument."

"Create the translate tool. It needs the caller's locale, so set forwardHeaders: ["x-locale"] so EgoX passes that header through from /ask."

"Create a knowledge-base entry titled RefundPolicy from the contents of docs/refund-policy.md."

"Show me every EgoX tool that came from MCP — I want to review them."

Good to know

Tokens are scoped and revocable. Every connection — browser sign-in or minted token — is scoped to one project and listed on the Console's MCP Tokens page, where you can revoke it at any time.
Least privilege. Token scopes (tools:read, tools:write, rag:read, rag:write) are enforced by EgoX — a read-only token can't modify anything, even if the agent tries.
Human review by default. MCP-created tools and docs are drafts (source = 'mcp') until you approve them, and human-curated content is protected from accidental overwrite.

Review agent-created tools & knowledge → Console
Integrate the resulting /ask API → Client SDK

Connect

What the agent can do

Tool-definition fields the agent can set

What you can ask the agent

Good to know

Related

On this page