Nemphi MCP server

Nemphi's MCP server is a remote Model Context Protocol server. It lets an AI assistant (Claude, ChatGPT, Cursor or any MCP client) read and work in your Nemphi workspace.

It is not a side door. Every call runs through the same pipeline as the app: your role's permissions, the workspace's autonomy policies, approvals and the action ledger. What you cannot do in the app, the assistant cannot do on your behalf.

Server URL: https://api.nemphi.com/mcp. Streamable HTTP transport, stateless: each request is one independent POST.

Claude (web, desktop and mobile). Open claude.ai, go to Settings › Connectors and look for Nemphi in the directory. If it is not listed yet, choose "Add custom connector" and paste https://api.nemphi.com/mcp. Claude sends you to sign in to Nemphi; pick the workspace and accept on the consent screen. Connectors are shared across the web, desktop and mobile apps.

Claude Code. One command:

claude mcp add --transport http nemphi https://api.nemphi.com/mcp

On first use, Claude Code opens the browser to sign in.

ChatGPT, Cursor and others. Any client that speaks Streamable HTTP and OAuth 2.1 with PKCE (S256) works. The server accepts dynamic client registration and Client ID Metadata Documents, so you do not need to ask us for a client id. Metadata is at https://api.nemphi.com/.well-known/oauth-protected-resource.

There are two ways to identify yourself.

OAuth. What connectors use. Two scopes: tools:read reaches read tools only; tools reaches read and write tools. Access tokens last 1 hour and refresh tokens 30 days; each refresh issues a new token and invalidates the old one.

API key. Create one in Nemphi under Developers › API keys and send it in the Authorization: Bearer nk_… header. The key inherits the role you assign it.

To remove access, disconnect the connector in the client or revoke the grant or the key from Nemphi. Either takes effect on the next call.

Nemphi has about 550 tools. Handing all of them to the model costs context and makes it pick worse, so the server exposes three tools and the assistant writes code that calls the rest.

Tool	What it does	Arguments
search_tools	Finds Nemphi tools by keyword. Read-only.	{ "query": string, "module"?: string, "limit"?: number }. limit is 1 to 20; default 8.
run_read_code	Runs a JavaScript snippet that may call read tools only.	{ "code": string }
run_code	Same, and may also call tools that change data. Every call still goes through the workspace's policies and approvals.	{ "code": string }

search_tools returns, for each match, the tool's name, title, description, risk class and JSON input schema.

Inside the snippet:

• nemphi.call(toolName, input) calls a tool synchronously and returns its JSON result. On refusal it throws an Error whose code property says why: approval_required, policy_denied, invalid_input, among others.
• Whatever you write with console.log(...) comes back in logs.
• The value of the snippet's last expression comes back in result. A bare return at the top level does not work; end with the expression you want returned.

Limits per run: 30 seconds, 64 MB of memory, 50 tool calls and a result trimmed to 100 KB. No network, filesystem, timers or import.

Example, with run_read_code:

const drafts = nemphi.call("invoices.list", { status: "draft" });

drafts.invoices.map((invoice) => ({
  number: invoice.number,
  total_cents: invoice.total_cents,
}));

Full catalog. A client that prefers one MCP tool per Nemphi tool can connect to https://api.nemphi.com/mcp?tools=all. Tool names are Nemphi's, for example invoices.list. Same pipeline; only the shape of the list changes.

Read tools run right away.

Tools that post, pay, publish or send outside Nemphi, and tools that change what agents may do on their own, follow the workspace's policies. When a call needs a person's approval it returns approval_required. The assistant should tell its user; the person decides in Nemphi's inbox and the assistant retries or moves on. An assistant never approves on its own.

Workspaces created by an agent. When an agent created the workspace and its API key through self-serve signup, its user rarely has the app open. In those workspaces a call that needs approval is not lost: it is saved as a pending approval, and the response carries the approval id and a short summary for the agent to show its user. The agent relays the user's answer with two MCP tools: approvals.pending lists pending approvals and approvals.decide approves or denies one. In code mode they are nemphi.call("approvals.pending", {}) and nemphi.call("approvals.decide", { … }) inside run_code. Nemphi then runs the tool in the owner's name and the ledger marks the action "approved via the assistant". The owner can still decide or reverse it from Nemphi's inbox, and revoking the key stops the relaying.

Purchases that move money (buying a domain, changing the plan or overage) are not available over MCP. Do them in the app.

Every call lands in the action ledger with the actor "MCP client", the name of the person who authorized the connection, the tool and the result.

• "Which invoices are overdue this month, and who do we need to chase?"
• "Create a draft invoice for Comercial Los Andes for 12,480,000 COP with the items from quote 231."
• "Summarize this quarter's expenses by category and compare them with last quarter."
• "Which tasks in the Bodega Norte project have no owner?"
• "Log in the CRM that I spoke with Paula at Distribuciones Andina today and that I call back on Tuesday."

401 or "unauthorized". The token expired or was revoked. Reconnect the connector, or create a new key if you were using one.

approval_required. The call is waiting for a person. Open Nemphi's inbox, decide, and ask the assistant to retry. In a workspace created by an agent, you can also answer the assistant and it relays your decision with approvals.decide.

Tool missing or policy_denied. Your role does not reach that tool, or the token was issued with tools:read and the tool writes. Ask an admin for the permission, or reconnect with the tools scope.

invalid_input. The assistant sent an argument the tool does not accept. The error message names it; search_tools returns the exact schema.

Nemphi receives only the tool calls the assistant makes: the tool's name and its arguments. It does not read your conversations or the rest of the assistant's context. What the assistant's provider does with the responses is governed by its own terms.

Data handling is described in the privacy policy. For questions or problems, write to [email protected].